# Changelog Product updates, model releases, and platform changes from the Brainbase team, newest first. RSS: https://brainbaselabs.com/changelog/rss.xml ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-11-10-fixes-and-improvements) 2026-11-10 The Usage page's Members, Agents, and Workflows tables now page through results instead of loading every row at once, Billing only appears in the sidebar once your account's access is actually confirmed, and sending feedback is one click from the sidebar. The Usage, Members, API keys, and Templates pages show your last results right away on a return visit, and phones get a back link in the top bar. Fixed mobile cards overflowing the screen, a command-palette crash from certain browser extensions, your project list and organization settings not clearing or reporting failure correctly when access changes, and an agent incorrectly saying it couldn't see an image after switching models mid-conversation. A custom Slack app can be set per agent to keep replying in a thread without another @mention, agents can stay quiet on Slack without a no-response message, and Meta's models now show Meta's official logo in the model picker. ### Console - The Usage page's Members, Agents, and Workflows tables now page through results 25 at a time instead of loading every row at once. - Fixed cards on the Deployments, home, templates, playbooks, and workflows pages overflowing off the right edge of the screen on narrow mobile widths. - Fixed the command palette (⌘K/Ctrl+K) crashing when a keydown event without a key — from a browser extension or password manager — reached it. - Template cards now show the agent's runtime as an icon and label, instead of a raw internal id. - Fixed a refused read of your project list leaving a stale list and admin controls on screen instead of clearing them, and an organization settings save by a non-owner admin now reports failure instead of silently saving nothing. - The Usage, API keys, and Templates pages show your last results right away when you come back to them instead of a spinner, and the Members page loads faster with its member lookups running in parallel. - On phones, the top bar has a back link to the page above, so you can get from an agent or orchestration page back to its project without using the browser's back button. - Moving a node on an orchestration canvas now retries the save if the connection drops, instead of reverting the canvas after one failed request. ### Billing - Billing in the sidebar no longer flashes into view and back out while your access is still loading — it now only appears once your actual billing access for that team is confirmed. ### Feedback - You can now send feedback — a bug, an idea, or anything else — from a new Feedback button in the sidebar. ### Slack - A custom Slack app can now be set per agent, from the agent's Surfaces page, to keep replying in a channel thread it's already working in without another @mention. It's off by default, the shared Slack app is unchanged, and top-level channel messages and replies in threads another bot owns still need a mention. - When an agent answers a Slack message only with a reaction, or intentionally stays quiet, the thread no longer gets an automatic no-response message. If an answer is actually lost, that message still appears. ### Models - Fixed an agent incorrectly telling you it couldn't see an image you sent after switching models mid-conversation; it's now told exactly which model it's running and whether that model accepts images. - Fixed images included in a tool's result being converted to garbled text instead of reaching certain models as an image, including Claude Fable 5.2 and Muse Ember. - Meta's models now show Meta's official logo in the model picker instead of a placeholder icon. ## [Gemini 4 Carbon Support](https://brainbaselabs.com/changelog#/r/2026-11-10-gemini-4-carbon-support) 2026-11-10 Google's Gemini 4 Carbon is now available on Kafka, OpenCode, Pi, and Qwen Code agents, with a 1M-token context window and image input, at introductory pricing through November 30, 2026. ### What's new - Gemini 4 Carbon is now available under All Models on Kafka, OpenCode, Pi, and Qwen Code agents. - It has a 1,048,576-token context window, up to 65,536 output tokens, selectable reasoning effort (low, medium, high), and image input. - Usage is billed at $2 per million input tokens and $10 per million output tokens. That introductory price runs through November 30, 2026, then rises to $4 and $20 on December 1, 2026. ## [Claude Fable 5.2 Support](https://brainbaselabs.com/changelog#/r/2026-11-09-claude-fable-5-2-support) 2026-11-09 Anthropic's Claude Fable 5.2 is now available on Kafka, OpenCode, and Pi agents, with a 1M-token context window and image input. ### What's new - Claude Fable 5.2 is now available under All Models on Kafka, OpenCode, and Pi agents. - It has a 1,000,000-token context window, up to 128,000 output tokens, selectable reasoning effort (low, medium, high, xhigh, max), and image input. - Usage is billed at $8 per million input tokens and $40 per million output tokens. ## [Muse Ember Support](https://brainbaselabs.com/changelog#/r/2026-11-09-muse-ember-support) 2026-11-09 Meta's Muse Ember is now available on Kafka, OpenCode, Pi, and Qwen Code agents, with a 1M-token context window and image input. ### What's new - Muse Ember is now available under All Models on Kafka, OpenCode, Pi, and Qwen Code agents: 1,048,576-token context window, up to 64,000 output tokens, selectable reasoning effort (low, medium, high), and image input. - $0.11 per million input tokens and $0.54 per million output tokens. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-11-08-fixes-and-improvements) 2026-11-08 Onboarding keeps your in-progress agent when you close the guide, now asks for a phone number first, and shows next steps after your first task finishes; first-session crashes are fixed, model pickers show each lab's real logo, and console, payments, sign-in, billing, and task reliability are improved. New accounts can now try a task before verifying a phone, the orchestration canvas gained undo and redo, a text-only task no longer fails because of a screenshot, Slack notifications are quieter and more reliable, and the Usage and Templates pages now load all at once instead of piece by piece. Canceling a payment setup or credit purchase no longer locks you out of trying again, switching conversations no longer leaves old messages on screen, and triggers and workflow tasks keep their original creator instead of being wrongly refused access to your connected integrations. GitHub review comments without a line reference are no longer dropped, and Slack replies are no longer dropped when Slack is slow to confirm a channel's history. ### Onboarding and agents - Closing the onboarding guide on the new-agent form no longer sends you back to the project dashboard and discards what you typed; it keeps you on the form and resumes at the right step afterward. - Fixed several first-session issues: a sign-in flow that could spin forever waiting on your session, a crash on pages viewed through a browser's auto-translation, a new-agent crash on devices without WebGL support, and a Memory tab error on a brand-new agent. - Onboarding's phone-verification step is now phone-first: enter a phone number to unlock a free model, see one clear next step afterward instead of a stale checklist, and the $25-credit card step moved out of onboarding into the sidebar and Billing, so new agents reach their first task sooner. - New accounts can now run one free task before verifying a phone number, so you see a result before being asked to verify. - The Brainbase starter templates shown during onboarding are now always available afterward too, on the Templates page and in New agent → From template. - The "What's next?" guide now appears after your first task finishes successfully, instead of popping up while it's still running and covering the result, with next steps for using a coding agent, building an orchestration, and connecting your tools. ### Orchestrations - Adding a trigger from the orchestration canvas no longer briefly flashes "project not found" before the page finishes loading. - Dragging an agent onto an orchestration now asks you to confirm the move instead of adding it immediately, so an accidental drop no longer silently pulls it out of your standalone agents. - The orchestration canvas now supports undo and redo for moving, connecting, adding, and removing agents and triggers. - Fixed the orchestration canvas unexpectedly zooming out after moving an agent or trigger that was still fully on screen. ### Models - Model pickers now show each lab's own logo — including DeepSeek, Particle AI, GLM, and Mistral AI — instead of a generic placeholder, and Gemini's logo now appears in its real colors. - Fixed incorrect context-window limits for GLM 5.2 and Kimi K2.7 Code on OpenCode and Pi agents, which could cause a long conversation to be rejected instead of automatically trimmed. - Fixed follow-up turns failing on Z.AI's GLM models, including the already-available GLM 5.3 FlashX, with an "Invalid API parameter" error. - Fixed a bug where switching models mid-conversation on Claude Code could leave the conversation unable to compact, making it permanently stuck. - GPT-5.6 Sol/Terra/Luna and GPT-6 Sol/Luna now reason on tool-calling turns in OpenCode and Pi agents, matching their behavior on other harnesses. - New agents and tasks that don't pick a model now default to GPT-6.1 Sol instead of GPT-6 Sol. - Fixed a bug where a screenshot captured by a tool call could fail an entire text-only task; it's now treated as a harmless limitation instead. ### Console - Fixed the top navigation bar cutting off organization, project, and agent names next to search on deep pages; names now stay whole and the search box shrinks first. - Fixed global search floating in the middle of empty space, and its results panel misaligning, when long breadcrumbs pushed it out of its centered position. - The Usage page now loads all at once instead of revealing the billing card, stat tiles, chart, and tables one at a time as each finishes loading. - The Templates page now shows your organization's templates together with the built-in Research, Plan, and Write starters in one list instead of as a separate section, and loads in one step instead of piece by piece. - On medium-width windows, opening a task now fills the page instead of squeezing the task list into a narrow strip, with a back button to return to the list. ### Payments - The Payments page now opens on a short intro explaining how agent payments work — your agent asks first, you approve each purchase in Link — instead of an empty wallet panel, for agents that have never connected a wallet. ### Sign-in and connect - Sign-in, connect, invite, and MFA pages now share one consistent step-by-step layout, and the connect page offers "Use another account" when you're signed in as the wrong one. - Fixed the emailed one-time code option disappearing from the sign-in page after switching to password sign-in and back. ### Billing - Fixed legacy Enterprise plans being locked out of paid models in the model picker even with a funded wallet. - Fixed canceling a payment method setup or a credit purchase part-way through sometimes locking you out of trying again for up to a day instead of letting you resume or restart it. ### Chat - Fixed message text drifting left as it wrapped to multiple lines in narrow chat columns, and fixed the placeholder text not returning after deleting everything you'd typed. - Fixed switching conversations sometimes leaving the previous conversation's messages on screen instead of clearing them before the new one loads. ### API - Fixed a task's creation time being sent without a timezone, which could make a client reading it in local time think the task started hours in the future. ### Slack - Status updates for an in-progress Slack message — queued, working, included in the reply — now show up as emoji reactions on your message instead of separate progress replies; a cancelled or failed message still gets a text reply so you know to resend it. - Fixed a bug where sending two DM messages to a coding agent in quick succession, before it replied, could make the second one wrongly show a "no reply" message even though the agent had already answered. - Fixed replies in a Slack thread your agent started itself sometimes never reaching the agent on a custom Slack app, instead of being treated as a new conversation. - Fixed some Slack replies being dropped, instead of just taking longer, when Slack was slow to confirm a channel's message history. ### Integrations - Fixed GitHub code review comments that don't include a line reference sometimes being dropped instead of reaching you. ### Reliability - Fixed an occasional error that could fail a message if the database connection dropped at the wrong moment. - Connecting to an MCP server is faster when the server is slow to start, and an agent's timeline now shows accurate connection status for it instead of reporting everything as healthy. - Fixed a case where one account's surge of task starts, or its backlog of recovery after a platform restart, could delay other accounts' new messages by minutes; starting tasks is now admitted fairly across accounts. - Error messages are clearer when a free or trial model isn't available right now, and a credits-exhausted message now mentions when a free model is available instead. - Free-model refusals on a non-standard sandbox now say exactly what to do — switch to the standard sandbox, or add credits — instead of advice that didn't fit your setup. - Fixed triggers, scheduled runs, and workflow- and orchestration-started tasks sometimes being refused access to your connected integrations, or blocked from billing, because the system lost track of who created them; they now keep their original creator from creation through execution. ## [Ripple-65B Preview Model Support](https://brainbaselabs.com/changelog#/r/2026-11-05-ripple-65b-model-support) 2026-11-05 Ricursive's Ripple-65B Preview is now available on Kafka, Codex, OpenCode, and Pi agents, with a 1M-token context window. ### What's new - Ripple-65B Preview is now available under All Models on Kafka, Codex, OpenCode, and Pi agents: 1,048,576-token context window and up to 131,072 output tokens. It doesn't offer selectable reasoning effort or image input on any harness. - $0.22 per million input tokens and $0.65 per million output tokens. ## [Faster First Replies in New Slack Threads](https://brainbaselabs.com/changelog#/r/2026-11-05-faster-slack-thread-replies) 2026-11-05 Starting a new conversation with your agent in Slack now gets it working noticeably sooner, since it starts on your first message right away instead of waiting on an extra readiness check first. ### What's improved - New Slack threads without file attachments now start the agent's first turn together with your first message, cutting out a round trip that used to make the agent wait before it even saw what you asked. ## [Slack DMs Now Stay Open Until You Reset Them](https://brainbaselabs.com/changelog#/r/2026-11-05-slack-dm-conversations-stay-open) 2026-11-05 Slack DM conversations with your agent no longer end after a few quiet hours, resetting one is now a single typed command, and top-level replies now work the same way across every kind of agent. ### What's new - A DM conversation now continues no matter how long it's been quiet, instead of automatically starting over after 4 idle hours; picking one back up after a couple of quiet hours shows a short reminder of when you last talked. - Type "!restart" in a DM, on any agent, to start a new conversation — replacing the New conversation button that used to show on every reply. - DM replies now land at the top level for every kind of agent, including coding agents, instead of only some of them, and agents installed as native Slack apps now treat DMs the same rolling-conversation way every other install does. ## [Members-Only Webhook Triggers for Orchestrations](https://brainbaselabs.com/changelog#/r/2026-11-04-members-only-webhook-triggers) 2026-11-04 Webhook triggers can now be restricted to people in your org instead of anyone who has the URL, with a single clear endpoint and a ready-to-copy example request, plus full management from the CLI and API. ### What's new - Every webhook trigger now has a "Who can call it" switch: Members, which requires an API key, or Anyone with the URL. New triggers default to members-only. - The trigger's page shows one endpoint that matches your choice, explained in plain language, plus a ready-to-copy example request for calling it. - New-trigger and trigger-detail pages now show guidance written for the trigger type you picked, instead of generic copy written for app triggers. - Webhook triggers can now be created, wired, switched between members-only and public, and fired directly from the CLI and the API, not just the canvas. ## [Tencent Hy4 Model Support](https://brainbaselabs.com/changelog#/r/2026-11-04-tencent-hy4-support) 2026-11-04 Tencent's Hy4 model is now available on Kafka, Codex, OpenCode, and Pi agents, with a 1M-token context window and image input. ### What's new - Hy4 is now available under All Models on Kafka, Codex, OpenCode, and Pi agents: 1,048,576-token context window, up to 65,536 output tokens, low/medium/high reasoning effort, and image input. It isn't offered on Claude Code or Qwen Code agents, where it couldn't read images. - $0.45 per million input tokens and $1.35 per million output tokens, Tencent's launch pricing through December 31, 2026. ## [Turn Any Agent into an Orchestration](https://brainbaselabs.com/changelog#/r/2026-11-04-agent-to-orchestration) 2026-11-04 Promote an existing agent into an orchestration in one step, from the project page or the CLI, instead of creating an empty orchestration and dragging the agent into it. ### What's new - On the project page, a "Make an orchestration" option on any standalone agent creates an orchestration named after it with that agent as its only member. - From the CLI, `brainbase orchestration from-agent` does the same thing from the command line. ## [Slack DMs Now Remember Context Outside Threads](https://brainbaselabs.com/changelog#/r/2026-11-04-slack-conversation-memory) 2026-11-04 Your agent's Slack bot now keeps a DM going as one conversation instead of starting over with every message, and a new @mention in a channel can see the bot's recent threads there. ### What's new - Direct messages are now one ongoing conversation: each new message carries the context of what came before, until the DM has been quiet for 4 hours or you press New conversation. - A new top-level @mention in a channel still starts a new task, but the agent now sees a short list of its own recent threads in that channel and can open one, so a follow-up like "also run it on the blog page" can find what "it" refers to. - Replies in a DM are posted at the top level instead of buried in a thread, matching how the conversation was asked. - Agents installed as native Slack agents keep one session per DM, as before. ## [Redesigned Chat and Agent Workspace](https://brainbaselabs.com/changelog#/r/2026-11-03-redesigned-workspace) 2026-11-03 Chat and the agent tabs got a visual refresh — tool calls collapse into readable summaries, browser replays get a real timeline, and empty tabs explain what they do and how to turn them on. ### Chat - A run of tool calls now collapses into a one-line summary you can open to see each call's input, with long output scrolling instead of pushing the page down. - Cleaner message bubbles and spacing, and a thinking-effort picker next to the model picker. - While a follow-up message is sending during a run, Stop is the only action shown until it's safe to send again. ### Browser - Session replays now have a scrubbable timeline with grouped markers, clip boundaries, and a 1x–16x speed picker. ### Agent tabs - Tasks, Benchmarks, Evaluations, Tools, Skills, Surfaces, Memory, Browser, Filesystem, Secrets, Playbooks, and Entrypoint each show a clear intro — what the tab does and how to turn it on — instead of an empty page before you've used them. - Memory, Browser, and Filesystem stay reachable from the sidebar even when off, with a switch right on the intro to turn them on. - The Analytics tab has been removed from the agent workspace. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-11-02-developer-updates) 2026-11-03 New per-item API routes for agent configuration, image-input info on the models endpoint, and more reliable MCP OAuth connections. ### API - New `/v2` routes change one agent setting at a time instead of resending a whole list: `PUT`/`DELETE /v2/agents/{id}/skills/{creator}/{slug}` for registry skills, `PUT`/`DELETE /v2/agents/{id}/mcp-servers/{name}` for MCP servers, `GET`/`PUT /v2/agents/{id}/playbooks` plus an archive route, and `PUT`/`DELETE /v2/orchestrations/{id}/members/{agent_id}`. `POST /v2/agents/{id}/evals/{eval_id}/run` starts a billable eval run. - `GET /v1/models` now reports `image_input` for every model, so you can tell which ones accept image input without checking the docs. ### MCP - Adding or updating an MCP server through the API or CLI now runs OAuth discovery automatically, so servers that need OAuth, such as Google's Gmail MCP server, can be connected without an extra step. - Signing in from the MCP consent page now happens in the same tab and returns you to finish connecting, instead of opening a new tab that could strand you partway through onboarding. - MCP clients like ChatGPT now get built-in guidance to suggest a first task right after creating an agent, clearer wording on when a run uses credits versus the free model, and the full message when a run is refused for billing or rate-limit reasons instead of a raw error. ## [Faster Follow-Up Replies Across Agents](https://brainbaselabs.com/changelog#/r/2026-11-02-faster-follow-up-replies) 2026-11-02 Replies now start sooner on every agent harness, since Brainbase overlaps backend checks that used to run one after another, and Claude Code agents keep their connection running between turns instead of restarting it. ### What's improved - On Claude Code agents, follow-up turns that involve billing, in-chat payments, or orchestration hand-offs now reuse the agent's already-running connection instead of restarting it, avoiding a multi-second delay that used to hit most of those turns. - Every model call now overlaps its authentication, conversation-state, and billing checks instead of running them one after another, shaving time off the start of every reply on every harness. - Claude Code follow-up turns on tasks started from the web app now also reuse their running connection instead of restarting it, closing a gap where only some follow-ups got this speed-up. - Sending a message or starting a task on a model that might be free no longer waits on its own for the free-model eligibility check, and that check itself is now faster too. - There's nothing to change: this applies automatically to existing agents and tasks. ## [New Model Support: GPT-6.1 Terra Pro, GLM 5.3 FlashX Legal, and More](https://brainbaselabs.com/changelog#/r/2026-11-02-new-model-support) 2026-11-02 Seven new models from OpenAI, Z.ai, MiniMax, Meta, StepFun, and NVIDIA are now available across the Brainbase model picker. ### What's new - GPT-6.1 Terra Pro, GLM 5.3 FlashX Legal, MiniMax M3.1 Flash, Muse Wonder, Step 5 Preview, Muse Spark Open, and Nemotron 3.5 Super VL are now available under All Models, with two harness limits noted below. - GPT-6.1 Terra Pro: 1,050,000-token context window, low/medium/high reasoning effort (always on), $1.08 per million input tokens and $5.40 per million output tokens. Available on Kafka, Codex, OpenCode, and Pi agents. - GLM 5.3 FlashX Legal: 1,048,576-token context window, low/medium/high reasoning effort, $0.40 per million input tokens and $1.35 per million output tokens. - MiniMax M3.1 Flash: 1,000,000-token context window, low/medium/high reasoning effort, $0.16 per million input tokens and $0.65 per million output tokens. - Muse Wonder: 1,048,576-token context window, low/medium/high reasoning effort, $4.32 per million input tokens and $21.60 per million output tokens. - Step 5 Preview: 1,000,000-token context window, low/medium/high reasoning effort, $1.08 per million input tokens and $2.92 per million output tokens. - Muse Spark Open: 1,048,576-token context window, $1.25 per million input tokens and $4.25 per million output tokens. - Nemotron 3.5 Super VL: 1,048,576-token context window, $0.20 per million input tokens and $0.50 per million output tokens. Available on Kafka, OpenCode, and Pi agents. - All seven support image input. ## [Mistral AI and GLM 5.3 Model Support](https://brainbaselabs.com/changelog#/r/2026-10-30-mistral-glm-5-3-support) 2026-10-30 Six new models from Mistral AI and Z.ai — including Mistral Large 4 and GLM 5.3 Prime — are now available across the Brainbase model picker. ### What's new - Mistral Large 4, Mistral Medium 3.5, Mistral Small 4, Devstral 2, GLM 5.3 Prime, and GLM 5.3 FlashX are now available under All Models in every harness picker. - Mistral Large 4: 1,048,576-token context window, low/medium/high reasoning effort, $0.76 per million input tokens and $2.27 per million output tokens. - Mistral Medium 3.5: 262,144-token context window, low/medium/high reasoning effort, $1.62 per million input tokens and $8.10 per million output tokens. - Mistral Small 4: 262,144-token context window, low/medium/high reasoning effort, $0.16 per million input tokens and $0.65 per million output tokens. It's available on Kafka, OpenCode, and Pi agents; it isn't offered on Claude Code or Codex agents, where it didn't reliably call tools. - Devstral 2: 262,144-token context window, $0.43 per million input tokens and $2.16 per million output tokens. - GLM 5.3 Prime: 1,000,000-token context window, low/medium/high reasoning effort, $3.02 per million input tokens and $9.50 per million output tokens. - GLM 5.3 FlashX: 1,048,576-token context window, low/medium/high reasoning effort, $0.40 per million input tokens and $1.35 per million output tokens. - Mistral Large 4, Mistral Medium 3.5, Mistral Small 4, and GLM 5.3 FlashX support image input. ## [New Accounts Get $25 in Free Credits Automatically](https://brainbaselabs.com/changelog#/r/2026-10-30-automatic-signup-credits) 2026-10-30 New accounts can now claim $25 in free credits by verifying a phone number and card, instead of waiting on manual approval. ### What's new - Verifying your phone number and a credit or debit card now grants $25 in free credits right away, instead of waiting on manual staff review. - Free credits run on a set of open models in the web app until your first top-up. A model your free credits can't run shows a "Top up" badge with a link to add credits, and new agents and tasks now default to a model the credits can run instead of one that's locked. - The claim now completes automatically once your phone, card, and first agent are all verified, whichever page you finish on, instead of requiring an extra click. - Onboarding no longer asks whether to build your first agent with the web app, API, or CLI; it takes everyone into the guided web setup, since the API and CLI unlock after your first top-up. ## [Webhook Triggers for Orchestrations](https://brainbaselabs.com/changelog#/r/2026-10-30-orchestration-webhook-triggers) 2026-10-30 Orchestrations can now be started by calling a webhook URL, alongside the existing app and schedule triggers. ### What's new - Adding a Webhook trigger to an orchestration gives it a unique URL; posting JSON to that URL starts a run, with no third-party app connection required. - Send an `Idempotency-Key` header on your request to avoid starting a duplicate run if you need to retry a call. - A webhook trigger can be turned on or off from the orchestration canvas, the same as app and schedule triggers. ## [GPT-6.1 Terra and Gemini 4 Argon Support](https://brainbaselabs.com/changelog#/r/2026-10-30-gpt-6-1-terra-gemini-4-argon-support) 2026-10-30 Two new models — GPT-6.1 Terra and Gemini 4 Argon — are now available in the Brainbase model picker. ### What's new - GPT-6.1 Terra is now available for Codex, OpenCode, and Pi agents, with effort settings from none to max. - Gemini 4 Argon is now available across agents and tasks, with low, medium, and high effort settings, at introductory pricing through 2026-11-30. - Both are featured in the model picker as new, primary models. ## [DeepSeek V4.1 Flash 1017 Support](https://brainbaselabs.com/changelog#/r/2026-10-30-deepseek-v4-1-flash-1017-support) 2026-10-30 DeepSeek V4.1 Flash 1017 is now available in the Brainbase model picker, with a 1M-token context window. ### What's new - DeepSeek V4.1 Flash 1017 is available under All Models in every harness picker, with text and image input, tool use, and reasoning. - Context window: 1,048,576 tokens. Max output: 262,144 tokens. - Pricing: $0.30 per million input tokens, $1.20 per million output tokens, $0.007 per million cached input tokens. ## [Particle-1 Model Support](https://brainbaselabs.com/changelog#/r/2026-10-30-particle-1-support) 2026-10-30 Particle AI's Particle-1 model is now available in the Brainbase model picker across every agent harness. ### What's new - Particle-1, from Particle AI, is offered in every harness picker, including Claude Code and Codex agents. - Pricing: $0.27 per million input tokens, $1.08 per million output tokens. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-10-30-developer-updates) 2026-10-30 Manage agent versions and tasks from the CLI and API, script the CLI more reliably, catch benchmark mistakes before a run starts, and connect MCP clients on their stable callback URL. ### Agent versions - `brainbase agent versions` lists an agent's versions newest first and marks the pinned one. `show` displays a version, and `diff` compares it with the one before it, another version, or the newest version. - `label`, `restore` (preview it with `--dry-run`), `pin`, and `unpin` work from the CLI too, and `brainbase task create --version v12` runs a task on a specific version. - The API adds `GET /v2/agents/{agent_id}/versions` and `GET /v2/agents/{agent_id}/versions/{version_id}`, with a `/diff` route for each version, and MCP clients get matching tools. From CLI v0.53.0, pin and unpin return a conflict if someone moved the pin in the meantime. ### Tasks - `brainbase task update` changes a task's title and labels. `--model` switches its model from the next turn, and `--clear-model` returns it to the agent's default. Over the API, `PATCH /v2/tasks/{task_id}` accepts `model`, `clear_model`, and `metadata_patch`. - Set labels when you create a task with `task create --metadata key=value`, and filter on them with `task list --metadata key=value`. `task get` now shows the task's model. - `brainbase task delete`, `agent delete`, and `orchestration delete` ask for confirmation, and scripts pass `--yes`. Deleting an agent deletes its tasks; deleting an orchestration keeps its member agents. ### Scripting - `brainbase orchestration init` writes a starter orchestration manifest, `brainbase --version` prints the installed version, mistyped commands get a suggestion, and `agent list --all-teams` lists every team at once. - With `--json`, most commands share one error shape and piped output drops spinners and color codes. Large output piped into tools like `jq` is no longer cut off at 64 KB when the CLI runs under Bun, for example through `bunx`. - `brainbase status` now reports any cloud change since your last sync, and `task logs --follow` waits for a queued follow-up instead of exiting. ### Benchmarks (beta) - A mistyped judge or variant model is refused when you plan a benchmark, with the model and evaluator named, instead of failing every trial at grading. Cursor, Factory, and Qoder agents are refused up front too, because those harnesses choose their own model. - Agent-judged evals on very long conversations now return pass or fail instead of a context-length error. ### MCP - Brainbase's MCP OAuth server now advertises issuer identification (RFC 9207) and returns `iss` on every authorize redirect, so connector clients like ChatGPT and Codex can use their stable callback URL instead of falling back to a per-server one. - A Pipedream copy-cell action no longer rejects valid cell partitions. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-10-29-fixes-and-improvements) 2026-10-29 Card payments add bank verification, sign-in and out-of-credit prompts say what to do next, and tasks, task Stop, triggers, project access, and Team upgrades behave more reliably. ### Billing - Buying credits or saving a card for auto top-up may now ask you to confirm the payment with your bank (3D Secure). If a card's security code doesn't match, the payment is refunded, no credits are added, and Billing explains why. - After three failed automatic payments in a row, auto top-up turns itself off, and Billing asks you to update your payment method. - Upgrading from Free to Team no longer risks canceling the subscription you just paid for, and clicking Upgrade while a Team checkout is open takes you back to that checkout. - When a task is refused because you're out of credits, a dialog now offers to get or add credits. - Sandbox prices are now in the docs. The Sandboxes page compares Daytona, E2B, Cloudflare, and Modal by size and hourly price, and the Sandbox Provider picker on the new-agent page links to it. - When staff approve a held signup credit grant, the account owner now gets an email saying so, instead of only finding out by reopening Billing. ### Sign-in - After a sign-in email is sent, the button counts down before offering to resend. An expired or already-used sign-in link says so and offers to send a new one, with your email filled in. - If phone verification can't text your country, or your number has opted out of texts, you now get a clear explanation instead of a generic prompt to try again. ### Tasks and triggers - Tasks that never started, for example because they were refused before setup, no longer show as running forever. They're closed with a note saying so, and they can be deleted. - On Claude Code agents, a follow-up sent while the first turn is still running now runs instead of being lost, and a retried task that succeeds no longer shows the earlier attempt's error. - Starting many tasks at once, for example from an orchestration, no longer fails some of them with a billing-unavailable error. - A webhook trigger that's switched off no longer starts agents, a schedule in Test mode no longer runs on its schedule, and the same webhook payload sent again after 24 hours starts a new run. Slow trigger tests keep checking and show how many agents have a task so far. ### Permissions and uploads - Connecting or disconnecting an app account on an agent now requires the agent's owner or a team admin. Other members see why connecting isn't available. - Agent icons must be PNG, JPEG, GIF, or WebP images. Public form uploads accept PDFs, images, Office documents, and TXT and CSV files, are checked against their real file type, and download instead of opening in the browser. Form submissions are rate-limited. ### Benchmarks - A benchmark bundle that fails to import now shows a plain-language error in the dialog instead of a raw API message. ### Security - Viewing an agent's related agents no longer reveals the names of agents in other projects you don't have access to. ### Chat - Pressing Stop while a task is starting up is now retried until confirmed, instead of leaving the agent running while the app said it had stopped. - If an agent actually finished a task after you pressed Stop, the task now shows that instead of offering to retry a message that already succeeded. ### Settings - Removed the "Display language" setting, which never actually changed anything. - [Sandbox pricing](https://docs.brainbaselabs.com/docs/agent/sandboxes) ## [New Agents Start on GPT-6 Sol, and Model Pickers Show Current Models](https://brainbaselabs.com/changelog#/r/2026-10-27-gpt-6-sol-default-current-models) 2026-10-27 Agents created without a model now start on GPT-6 Sol, or Claude Opus 5.5 on Claude Code, and model pickers list current models newest first. ### What's new - New agents and tasks that don't name a model now start on GPT-6 Sol, or Claude Opus 5.5 on Claude Code agents. Cursor, Factory, and Qoder agents keep choosing their own model. - The same default applies whether you create an agent in the console, with `brainbase agent create`, or through the API. `GET /v2/models` lists the default for each harness. - Model pickers now list models newest first and hide models released before July. Agents and tasks already on one of those models keep it. - Agents that already have a model keep it. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-10-24-developer-updates) 2026-10-24 Set up surfaces, OAuth MCP servers, and integration apps from the terminal, and keep the CLI signed in separately from the browser. ### Connect from the CLI - `brainbase agent surface` shows and changes the settings of an agent's surfaces, including Chat, Slack, Meetings, GitHub, and Linear. Steps that need a provider's consent open a browser link, and the CLI waits for the result. - `brainbase mcp connect ` authorizes an OAuth MCP server through a one-time browser link, and `mcp list` no longer reports working OAuth servers as expired an hour after you connect them. Add `--agent ` to run either from any folder, without `brainbase link`. - `brainbase integrations` manages a Kafka agent's integration apps: search the catalog, add an app, and connect an account. ### Sign-in - `brainbase login` now creates a session of its own, so the CLI and the web app no longer sign each other out. Run `brainbase login` once to switch over. - When a session can't be renewed, the CLI says why and what to run instead. If you've saved a personal access token for the same account with `brainbase token create`, commands fall back to it automatically and print a one-line notice. ## [Kimi K3.1 Pro Support](https://brainbaselabs.com/changelog#/r/2026-10-23-kimi-k3-1-pro-support) 2026-10-23 Kimi K3.1 Pro, Moonshot AI's newest Kimi model, is now available in Brainbase. ### What's new - Kimi K3.1 Pro is marked New in model pickers wherever Kimi K3 is offered, including Kafka, Claude Code, Codex, and OpenCode agents. Kimi K3 remains available. - It has a 1M-token context window and accepts image input. - Existing agents keep their current model unless you switch. - Usage is billed at $2.16 per million input tokens and $10.80 per million output tokens. ## [New Tasks Start Faster](https://brainbaselabs.com/changelog#/r/2026-10-22-faster-task-starts) 2026-10-22 The first reply on a new task now arrives several seconds sooner on OpenCode, Claude Code, and Kafka agents. ### What's improved - Median time from your first message to the agent's first output dropped from about 19s to 16s on OpenCode, 24s to 19s on Claude Code, and 25s to 20s on Kafka. - Brainbase now creates the sandbox and warms up the agent while your first message is still being checked, instead of one step after another. - Kafka agents connect to all their MCP servers at once instead of one at a time. - There's nothing to change: new tasks get this automatically. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-10-22-fixes-and-improvements) 2026-10-22 Kafka agents end their turn with a real reply, follow-ups send right away, and your own messages take up less space in the thread. ### Agents - A Kafka agent's turn now ends when it replies, so the thread shows its answer instead of an empty "Task completed". Agents still stop and ask when they need something from you. - This is on for every Kafka agent. New sandboxes get it right away, and existing threads pick it up when their sandbox restarts. To keep the previous behavior for one agent, set the agent secret `BRAINBASE_END_TURN_ON_REPLY` to `0`. ### Chat - Follow-ups post immediately, and "Sending…" clears as soon as Brainbase accepts the message. - Your messages now appear as a compact right-aligned bubble with the time on hover, and the thread no longer jumps when a message is saved. - The agent name in the chat sidebar links back to the agent, and the task page's First token stat now measures the first streamed output instead of the end of the turn. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-10-09-fixes-and-improvements) 2026-10-09 Meeting bots get a screen-sharing setting, GPT-5.6 Luna and GPT-5 requests with custom sampling no longer fail, and chat and sign-in are smoother. ### Meetings - A "Let the bot see the screen" toggle on the Meetings card sets whether, by default, the bot can look at what's shared on screen or camera during a call. New meeting installations start with it on, and existing ones keep their current behavior until you change it. ### Models - Requests to GPT-5.6 Luna, GPT-5, and GPT-5 Mini no longer fail when a temperature or top-p value is set. Brainbase omits those settings for these three models, so they run with their own defaults. ### Chat and sign-in - Starting a new task keeps your line breaks, and long unbroken text like URLs or tokens wraps inside the message instead of overflowing it. - Entering your emailed code after you've already opened the sign-in link now signs you in instead of showing an error. - Thread timestamps and date search use your local time zone, so dates no longer appear a day off. - Expired attachment images refresh automatically instead of showing a broken thumbnail. ## [Claude Haiku 5.5 Support](https://brainbaselabs.com/changelog#/r/2026-10-07-claude-haiku-5-5-support) 2026-10-07 Claude Haiku 5.5, Anthropic's fastest and lowest-cost Claude model, is now available in Brainbase. ### What's new - Claude Haiku 5.5 is featured across compatible model pickers. Agents already using Claude Haiku 4.5 keep it. - It has a 1M-token context window, up to 128K output tokens, and adaptive-thinking effort controls from low through max. - Existing agents keep their current model unless you switch. - Usage is billed at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens. ## [Payments: Team Requests From Slack and Clearer Receipts](https://brainbaselabs.com/changelog#/r/2026-10-02-payments-update) 2026-10-02 · Beta Teammates can ask an agent for a Link purchase from Slack, payment receipts are easier to read, and approvals reach the agent in seconds. ### What's new - A "Team requests from Slack" switch on the Payments page lets the wallet owner allow Slack-started conversations with an agent to request a Link purchase from the owner's connected wallet. The wallet owner still approves every payment in Link. - Payments results and approvals now also reach agents running on Cloudflare and E2B sandboxes. ### What's improved - Each payment card shows the merchant, amount, and status once, with a plain-language summary, and timestamps are shown in your local time. Disconnecting your Link wallet now asks you to confirm first. - Approving or denying a Payments request in Link reaches your agent's task within seconds instead of several minutes, and an agent with an open request no longer says the payment failed while your decision is pending. - If disconnecting your Link wallet hasn't finished yet, the wallet panel offers Connect alongside Retry disconnect, so you're no longer stuck until the old disconnect clears. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-10-02-fixes-and-improvements) 2026-10-02 Owners can let a trusted admin manage other admins, long OpenCode runs are no longer canceled while a subagent is still working, and Slack stop controls, the Payments page, and pasted playbook content behave more reliably. ### Agents - OpenCode agents that hand work to a long-running subagent are no longer canceled after about an hour while the subagent is still working. Runs that are truly stalled are still detected and stopped. - Several small fixes make the task activity view clearer and easier to follow. ### Teams - Owners can turn on "Admin manager" for an admin from that admin's role menu on the Members page, letting them change the role of or remove other admins. Only the owner can grant or revoke it, and billing and other owner-only settings are unaffected. ### Integrations - Stopping a task from Slack now gives immediate, accurate feedback, whether the stop succeeds or is denied. ### Payments (beta) - The agent Payments page keeps working even when part of the underlying data is malformed or temporarily unavailable, instead of failing to load. ### Playbooks - Pasting a playbook's entire contents into a brand-new, empty playbook — or navigating away moments after editing — could silently fail to save. Both cases are now fixed. ### Platform - Opening a task from a Slack notification link could land you in an outdated chat view or the wrong project; those links now resolve to the correct task directly. ## [GPT-6.1 Sol Support](https://brainbaselabs.com/changelog#/r/2026-09-29-gpt-6-1-sol-support) 2026-09-29 OpenAI's GPT-6.1 Sol is now available for Brainbase agents using Kafka, Codex, or OpenCode, at the same input and output rates as GPT-6 Sol. ### What's new - GPT-6.1 Sol is available and marked New in the model pickers for Kafka, Codex, and OpenCode agents, and is now featured across all three, including Codex's Featured pane and the main model dropdown. - Start a new task to use GPT-6.1 Sol if your current sandbox predates this release. - It supports text and image input, a 1,050,000-token context window, up to 128,000 output tokens, and low, medium, high, xhigh, and max reasoning effort. - GPT-6.1 Sol uses Standard execution; Fast is not available. - GPT-5.6 Sol remains the default, so existing agents keep their current behavior unless you select GPT-6.1 Sol. - Kafka Legacy and the other agent harnesses are not included in this release. ### Pricing - GPT-6.1 Sol is billed at $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. - Requests above 272K input tokens use OpenAI's long-context rates. - [OpenAI's GPT-6.1 Sol announcement](https://openai.com/index/introducing-gpt-6-1-sol/) - [GPT-6.1 Sol API model details](https://developers.openai.com/api/docs/models/gpt-6.1-sol) ## [Claude Sonnet 5.5 Support](https://brainbaselabs.com/changelog#/r/2026-09-28-claude-sonnet-5-5-support) 2026-09-28 Claude Sonnet 5.5, Anthropic's latest Sonnet model with the best combination of speed and intelligence, is now available in Brainbase. ### What's new - Claude Sonnet 5.5 is featured and marked New across compatible model pickers, placed ahead of Claude Sonnet 5. Existing Claude models remain available. - It has a 1M-token context window, up to 128K output tokens, and adaptive-thinking effort controls from low through max. - Claude Code keeps Claude Opus 5 as its default, so existing agents keep their current behavior unless you switch models. - Usage is billed at $2 per million input tokens and $10 per million output tokens. ## [Payments (Beta) Now Open to Every Account, With Higher Limits](https://brainbaselabs.com/changelog#/r/2026-09-28-payments-open-to-every-account) 2026-09-28 · Beta Payments — our beta wallet-approval flow for agent purchases — is now available to every account instead of a pilot group, supports orders above the previous $5 cap, and gives you clearer status and a way to recover a stuck request. ### What's new - Payments is no longer limited to a pilot group — it now appears in the sidebar for every signed-in account. - The previous $5 pilot limit on agent purchases has been removed, so larger orders are no longer rejected. - The Payments page now shows shipping, fees, and discounts as part of an order total, not just the subtotal. - If a payment request ends up in an unclear state, you can now recover it yourself with a new Recover request action instead of it staying stuck. ### What's improved - Status wording is clearer about what's agent-reported versus independently confirmed, Cancel is available in more request states, and owners can mark a request resolved. - Payment reads and cancellations now retry automatically through brief backend contention instead of surfacing an error on the first hiccup. ## [Cloudflare Sandboxes Now Available to Every Account](https://brainbaselabs.com/changelog#/r/2026-09-28-cloudflare-sandboxes-open-to-everyone) 2026-09-28 Cloudflare-backed sandboxes, a faster agent runtime option, are no longer limited to internal accounts and now show by default when starting a new agent, with more reliable file delivery to a task that's gone idle. ### What's new - Cloudflare is now a default option in the new-agent picker for every account, not just internal ones. - Sending a file to a task whose Cloudflare sandbox has gone idle now reliably wakes it and delivers the attachment, instead of the handoff timing out. ## [Live Progress, Questions, and Stop Controls for Agents in Slack (Beta)](https://brainbaselabs.com/changelog#/r/2026-09-28-slack-agent-progress-approvals-stop) 2026-09-28 · Beta Agents connected to Slack can now post live progress updates, ask clarifying questions, and be stopped mid-task right from the thread, with an optional approval step before sensitive tool actions that's off by default. ### What's new - Agents running through Slack now post concise, live progress updates and can ask a clarifying question directly in the thread, with a link back to the full task. - You, or an eligible teammate, can stop a running task from Slack; by default only the person who started it can stop it. - Workspaces can optionally turn on an approval step that pauses a sensitive tool action in Slack until someone approves it — off by default. ### What's improved - Slack no longer shows duplicate, indistinguishable activity labels for the same task step, and a rejected Stop now gets a clear explanation instead of no feedback. - Fixed a case where a Stop request could be lost, leaving a task's Slack card incorrectly showing an error. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-09-25-fixes-and-improvements) 2026-09-25 GPT-5.6 agents can search the web live, Payments tracks purchases and merchant outcomes, team membership changes are safer, and task runs, live updates, and message sending are more reliable. ### Agents - Agents running GPT-5.6 models now have a live web-search tool built in, instead of falling back to a slower browser-based workaround to look things up. - Agent tasks now recover more reliably when the platform restarts mid-run, instead of occasionally getting stuck or colliding with a duplicate resume. - Agent tasks recover more reliably from brief connection interruptions and model-routing races, so runs are less likely to be misreported or repeat a step. - The console now automatically catches back up on task and thread updates after a dropped connection, a backgrounded tab, or a brief outage, instead of needing a page reload. ### Account and Billing - When a task fails because your account is out of credits, it now reports that clearly instead of showing a generic failure, and our docs explain exactly how to recover. - Accounts approaching their credit limit no longer receive the same usage-alert email repeatedly; it's now sent once per threshold crossing. - Grok 4.7 now appears promoted near the top of model pickers, backed by expanded routing and reasoning support, with platform LLM credits discounted 75% through the end of November. - Enterprise accounts now see their exact billing period and next invoice date on the Billing page, plus a simpler account picker when there's only one account to manage. ### Security - We patched a critical remote-code-execution vulnerability in image handling on our documentation site, and applied the same precautionary security updates across backend services as a hardening measure. No customer data was affected and no action is required. - Removing someone from a project is safer now: they're no longer silently re-added to the General group on login, and admins can't accidentally strip a member's last project. - Members without admin access now see owners, admins, and the teammates who share a group with them, instead of the entire organization roster, and the list is sorted by role. ### Chat - Chat attachments and follow-up messages now give clearer, less cluttered feedback, so it's easier to tell what's happening with an upload or a message you sent while a task was already running. - Sending a chat message during a brief network interruption no longer risks submitting it twice, and a temporary connection hiccup no longer makes your workspace list falsely appear empty. ### Payments (beta) - The Payments page now shows what your agent bought through Link — including delivery details and whether the merchant order actually went through — so you can review and, if needed, recover a purchase in one place. This continues our beta wallet-approval rollout. ### Platform - A settings change made in one browser tab could get silently overwritten if an older tab reloaded its own cached preferences afterward — including your instant Slack-reply setting. That's now fixed. ## [Wallet-Connected Purchase Approvals for Agents](https://brainbaselabs.com/changelog#/r/2026-09-24-wallet-approvals-for-agent-purchases) 2026-09-24 · Beta You can now connect a payment wallet and approve or decline an agent's purchase requests in real time, with the outcome delivered straight back to the task. This is an early rollout for test approvals only — no purchase is completed yet. ### What's new - Connect a payment wallet on the agent Payments page, then review and approve or decline a purchase request an agent is asking to make. - Your approval or decline is delivered back to the agent's task automatically, so the run continues without you needing to check back manually. - This is a limited early rollout: approvals are test-only right now and don't execute a real purchase. ### What's fixed - Fixed an issue in the hosted approval flow where a local-only test button could appear and produce a confusing error after you'd already approved the request. - Fixed a bug where your payment status could briefly show as missing right after approving or continuing a request. ## [GPT-6 Sol and GPT-6 Luna Support](https://brainbaselabs.com/changelog#/r/2026-09-22-gpt-6-sol-luna-support) 2026-09-22 OpenAI's GPT-6 Sol and GPT-6 Luna are now available in Brainbase, with standard token rates at least 50% lower than their GPT-5.6 counterparts. ### What's new - GPT-6 Sol and GPT-6 Luna are featured and marked New across compatible model pickers, next to GPT-6 Astra. The GPT-5.6 models remain available. - GPT-6 Sol is built for complex reasoning and coding; GPT-6 Luna is the cost-efficient choice for high-volume, well-defined tasks. Both support text and image input, a 1,050,000-token context window, up to 128,000 output tokens, and reasoning effort from none through max, plus Fast mode. - GPT-5.6 Sol remains the default, so existing agents keep their current behavior unless you switch models. ### Pricing - GPT-6 Sol is billed at $2 per million input tokens, $0.20 per million cached input tokens, and $10 per million output tokens, half of GPT-5.6 Sol. - GPT-6 Luna is billed at $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.50 per million output tokens. - Requests above 272K input tokens use OpenAI's long-context rates, and Fast mode costs twice the applicable rate. - [OpenAI's GPT-6 Sol and Luna announcement](https://openai.com/index/introducing-gpt-6-sol-and-luna/) - [GPT-6 Sol API model details](https://developers.openai.com/api/docs/models/gpt-6-sol) - [GPT-6 Luna API model details](https://developers.openai.com/api/docs/models/gpt-6-luna) ## [Claude Opus 5.5 Support](https://brainbaselabs.com/changelog#/r/2026-09-22-claude-opus-5-5-support) 2026-09-22 Claude Opus 5.5, Anthropic's latest Opus model for long-running agentic coding and knowledge work, is now available in Brainbase at a lower price than Claude Opus 5. ### What's new - Claude Opus 5.5 is featured and marked New across compatible model pickers, placed ahead of Claude Fable 5.1 and Claude Opus 5. Existing Claude models remain available. - It supports adaptive-thinking effort controls from low through max, with a 1M-token context window and up to 128K tokens of output. - Claude Code keeps Claude Opus 5 as its default, so existing agents keep their current behavior unless you switch models. - Usage is billed at $4 per million input tokens and $20 per million output tokens. ## [Steer Running Agent Tasks with Follow-Ups and Stop](https://brainbaselabs.com/changelog#/r/2026-09-22-agent-task-followups-and-stop) 2026-09-22 You can now send follow-up instructions to an agent task while it's still running, or stop it cleanly, instead of waiting for the run to finish or starting a new one. ### What's new - Send a follow-up message to a task that's already in progress — the agent picks it up mid-run instead of ignoring it or requiring a brand-new task. - Stop a running task explicitly from the CLI (v0.35.0) or a connected messaging integration, with the agent wrapping up its current step cleanly instead of being killed mid-action. - Follow-ups and stop requests work the same way whether a task was started from chat, a workflow connector, or the CLI. ## [Grok 4.7 Support](https://brainbaselabs.com/changelog#/r/2026-09-21-grok-4-7-support) 2026-09-21 Grok 4.7 brings xAI's latest reasoning, coding, and image understanding to Brainbase, with a 500K-token context window and an extra-high reasoning option. ### What's new - Grok 4.7 appears near the top of compatible model pickers with a New badge. Existing Grok models remain available. - Choose low, medium, high, or extra-high reasoning effort; high is the default. Text, image input, tool calling, and structured outputs are supported. - Standard pricing is $2 per million input tokens and $6 per million output tokens. Prompts of 200K tokens or more use $4/$12 rates; cached input costs $0.50/$1 per million respectively. - [xAI Grok 4.7](https://docs.x.ai/developers/grok-4-7) ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-09-18-fixes-and-improvements) 2026-09-18 Codex agents can reach the network, free signup credits get a review step, and agent runs, the Stop button, and Slack file delivery are more reliable. ### Agents - Agent runs are less likely to fail right when they start, whether they're waiting on a connected tool to finish loading or waking a paused sandbox back up. - Agent conversations and managed tasks now recover automatically from occasional model-provider glitches instead of dying part-way through or breaking the rest of the conversation. - Project and orchestration cards now let you preview who's a member right from the list, without opening each one individually. - Agent tasks are now less likely to fail outright when the underlying model drops a response mid-stream or returns a malformed tool call — Brainbase now recovers automatically instead of ending the task. - Tasks running on the Codex agent runtime can now make outbound network requests, like fetching a webpage, while keeping their existing filesystem restrictions. - Tasks that delegate work to a sub-agent on the OpenCode runtime no longer stall for nearly an hour before failing — delegated work now completes normally. ### Account and Billing - Tasks that can't proceed because an account is out of credits now consistently show a clear billing message instead of a confusing generic failure. - Billing and Usage pages now show the details that actually apply to your plan, instead of mixing in concepts from the other billing model. - Billing and Usage pages for Enterprise accounts now reflect live wallet and invoice data instead of stale counters or a hardcoded credit price. ### Integrations - Files an agent sends to Slack now travel through the same durable delivery pipeline used elsewhere, so uploads no longer silently fail to reach the thread. - When an agent handed part of a Slack request off to a delegated task, the result could silently fail to come back; delegated work now reliably finds its way back to the original task and thread. ### Security - New email/password signups now go through a brief manual review before free trial credits are applied, to help prevent abuse. ### Playbooks - Opening a playbook right after switching to a different agent no longer shows a false "Playbook not found" error. ### Chat - Pressing Stop now reliably cancels the agent's response, including right after you send a message or right as a reply finishes. ## [Attach and Download Files Up to 1 GB](https://brainbaselabs.com/changelog#/r/2026-09-16-1gb-file-transfers) 2026-09-16 Chat attachments and task files now support uploads and downloads up to 1 GB, and an interrupted upload automatically resumes instead of starting over. ### What's new - Chat attachments and task Files now support files up to 1 GB, up from the previous 25 MB limit. - Uploads resume automatically after a dropped connection or a page reload, and a canceled or interrupted upload can be retried for 24 hours before it expires. - Downloads no longer buffer the entire file in your browser, so large files transfer reliably without stalling the tab. ### Good to know - Files placed in a workspace's Shared folder keep the existing 25 MB upload / 50 MB download limits. - Signing out clears any in-progress upload recovery data from your browser. ## [GPT-4o Retired, Existing Selections Move to GPT-5.6 Luna](https://brainbaselabs.com/changelog#/r/2026-09-11-gpt-4o-retired) 2026-09-11 GPT-4o and GPT-4o Mini are no longer offered anywhere in Brainbase — agents, workflows, and API/CLI requests that referenced them now run on GPT-5.6 Luna automatically, with no action needed on your end. ### What's changed - GPT-4o and GPT-4o Mini have been removed from model pickers across agents, workflows, and the phone integration. - Any saved agent, workflow, or API/CLI request that referenced GPT-4o or GPT-4o Mini — including dated snapshots — now runs as GPT-5.6 Luna automatically. - Luna has its own behavior and pricing; historical usage and billing records from before the switch are unchanged. ### Also fixed - Agent cards and overview pages now correctly display the replacement model instead of continuing to show GPT-4o after the migration. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-09-11-fixes-and-improvements) 2026-09-11 Benchmark runs can pause and resume with upfront cost estimates, model evaluations are billed like standard usage, custom MCP connections are hardened, and Slack, Linear, and file attachments are more reliable. ### Integrations - Agent replies delivered to Slack now stay correctly bound to the message that started them, even after a retry — preventing replies from arriving against the wrong request. - Adding a no-login app like Playwright to a brand-new agent could fail with an account-lookup error — that's now fixed, with explicit per-agent add and remove. - Stopping a Linear-triggered run now actually cancels it, task status and timestamps are accurate, and messages are more reliably attributed to their real source. - A new opt-in setting for the GitHub integration makes sure a review your agent claims to have finished actually gets posted, retrying once and withdrawing its acknowledgment if it still doesn't. - Custom Slack app integrations no longer start extra agent turns or show a misleading “no response was delivered” message when the agent's own reply happens to mention the app. ### Agents - Attaching a file to a new task no longer gets stuck retrying or times out — you now see a clear message right away when a file can't be attached, and stopping or switching tasks reliably cancels an upload in progress. - The add menu on an agent's own canvas no longer offers to create things that page can't actually create. ### Platform - Gemini tool calls and their results now reliably survive to the model's next turn, instead of silently breaking multi-step tool use. - If an MCP tool server your agent is connected to returns a server error or drops its connection mid-call, the run now fails that one tool call within milliseconds and reconnects for the next call, instead of hanging for up to an hour. ### Benchmarks (beta) - A benchmark run that hits a credit or plan limit now pauses instead of scoring the blocked case as a failure, and you can resume it right where it left off. Runs also show an upfront sandbox-time cost estimate, and can now fan out to more concurrent case runs (beta). ### Security - Custom MCP server connections are now restricted to public HTTPS destinations, and OAuth credentials are now bound to the exact server they were issued for — closing gaps that could let a malicious or misconfigured server redirect traffic or intercept authorization. ### Account and Billing - Evaluations that grade agent responses (used by benchmarks and quality monitoring) now consume credits the same way as a regular model call, instead of running for free. ### CLI and API - Streamed agent responses no longer come back silently empty or truncated when the upstream connection ends early. ## [Linear Surface Is Now Generally Available](https://brainbaselabs.com/changelog#/r/2026-09-10-linear-general-availability) 2026-09-10 Linear is now a fully supported Surface, open to every agent: connect it like Slack, and your agent can be delegated issues or @mentioned right in Linear, then work and reply there as itself. ### What's new - Linear is now generally available as a communication Surface for every agent — no beta gate, no requesting access. - Connecting Linear installs the agent as its own user in your workspace, the same way Slack does: teammates can delegate an issue to it or @mention it in a comment, and it shows up in Linear as a real teammate rather than a bot integration bolted on the side. - The agent picks up delegated or mentioned issues, posts a running log of its progress in the issue, and delivers its final answer there — asking a follow-up question in the same thread if it needs one. - It can also read and search issues, comment with real @mentions, move issues between workflow states, and update fields like title, description, priority, assignee, and labels, all attributed to it by name in Linear's history. ### Also in this release - Reconnecting an agent's Linear setup now defaults to reusing the existing app: it shows the previously connected workspace and prefills the app's client ID, instead of walking you back through first-time app creation. - An unfinished OAuth setup is now called out separately from a fully connected app, so it's clear what step you're on. - The connected-surfaces count in the agent sidebar now correctly includes GitHub and Linear alongside Slack and Meetings. ## [GPT-6 Astra for Kafka Agents](https://brainbaselabs.com/changelog#/r/2026-09-04-gpt-6-astra-kafka-support) 2026-09-04 Kafka agents can now use GPT-6 Astra for text and image tasks, with reasoning controls from low through max. ### What's new - GPT-6 Astra is now available for Kafka agents, featured and marked New. - Start a new task to use Astra if your current Kafka sandbox predates this release. - Use text and image input with low, medium, high, xhigh, or max reasoning effort. - Astra uses Standard execution; Fast is not available. - GPT-5.6 Sol remains the default. Select Astra when you want to use it; existing agents keep their selected model. - Kafka Legacy is not included in this release. - [Supported models](https://docs.brainbaselabs.com/docs/models) ## [GPT-6 Astra for Codex and OpenCode Agents](https://brainbaselabs.com/changelog#/r/2026-09-04-gpt-6-astra-support) 2026-09-04 GPT-6 Astra is now featured for Brainbase agents using Codex or OpenCode, with text and image input and reasoning controls from low through max. ### What's new - GPT-6 Astra is featured and marked New across compatible model pickers for Codex and OpenCode agents. - It supports text and image input, up to 128,000 output tokens, and low, medium, high, xhigh, and max reasoning effort. - The API supports up to 1,050,000 context tokens. Codex defaults to 272,000 and can be configured up to 872,000. - GPT-5.6 Sol remains the default, so existing agents keep their current behavior unless you select Astra. - Launch support is limited to Codex and OpenCode; Kafka support is deferred. ### Pricing - Standard short-context usage is $10 per million input tokens, $1 per million cached input tokens, $12.50 per million cache-write tokens, and $50 per million output tokens. - Requests above 272K input tokens use OpenAI's long-context rates. Batch and Flex cost 50% of Standard, while Fast costs twice the applicable rate where available. - [OpenAI's GPT-6 Astra announcement](https://openai.com/index/gpt-6-astra/) - [GPT-6 Astra API model details](https://developers.openai.com/api/docs/models/gpt-6-astra) ## [AWS Marketplace Billing Support](https://brainbaselabs.com/changelog#/r/2026-09-04-aws-marketplace-billing) 2026-09-04 Kafka can now be deployed as an AWS Marketplace appliance with billing handled entirely by AWS, so those customers manage their subscription without ever seeing Stripe. ### What's new - Customers who deploy Kafka through AWS Marketplace get a dedicated Subscription panel that reflects their AWS entitlement status instead of Stripe billing, plans, and credits screens. - Plan limits and team-invite restrictions no longer block these deployments, since billing and entitlement are managed entirely through AWS. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-09-04-fixes-and-improvements) 2026-09-04 Benchmarks get editable rubrics and trend charts, agent-judge evals run on every runtime, schedule triggers pause clearly when credits run out, and browser proxy passwords get stronger encryption. ### Agents - Per-agent task history and its overview totals could previously miss or undercount tasks beyond the first 50; both now reflect your full selected date range. - Agent-judge evals previously produced no verdicts at all for Claude Code, Codex, and OpenCode agents; they now run and report results just like they already did for Kafka agents. - Live task views now pick back up where they left off after a brief connection drop, instead of replaying the whole run from the start. - Clicking a task under Recent tasks on the agent overview now takes you straight to that task's detail panel, instead of dropping you on the general tasks list. - Orchestration memory now has its own row in the sidebar, so turning on shared memory for an orchestration no longer requires digging into Settings. ### Account and Billing - A recurring schedule trigger that can't afford its next run now pauses itself instead of retrying forever, and tells you exactly why. - Inkling's free early-access period has ended. Usage is billed at standard per-token rates, and the model picker now shows accurate pricing instead of $0. ### Benchmarks (beta) - Agent Benchmarking (beta) now gives each test case a real Checks section with a structured rubric editor, and shows exactly which criteria the judge found met or missed. - The benchmarks overview now charts every comparable run in order, so you can see how a benchmark trends across runs and drill into any single result without leaving the chart. ### Security - Proxy passwords you save for agent browsing are encrypted at rest and decrypted only at the moment they're used to route a request. - Fixed a security issue where a maliciously crafted package could cause the CLI to extract files beyond what its safety checks verified. Update to the latest CLI to get the fix. ### Integrations - Mentioning your agent on a pull request while it's already working now queues the request instead of failing with a confusing configuration error. - Coding tasks triggered from Slack now recover cleanly from brief failures instead of silently dropping or re-running the task. ### Workflows - Slack macro deliveries no longer get retired because of brief, unrelated network blips, so scheduled and triggered macros reach Slack more reliably. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-09-03-developer-updates) 2026-09-03 New CLI commands list, inspect, and wait on tasks, scaffold agents, and manage benchmarks and evals, with safer defaults for destructive and automated commands. ### CLI - If a `brainbase` command is interrupted mid-login, later commands now report a clear, actionable timeout instead of crashing with a raw stack trace. - You can now manage tasks straight from the terminal — list them, check status, tail logs, and block a script until one finishes — without opening the web app. - The CLI no longer silently confirms destructive actions when run without a terminal, and won't push local changes over cloud data it has never seen. - Benchmarks can now be authored, imported, run, and inspected without leaving the terminal — the CLI covers the full lifecycle from draft to results. - The CLI can now create a starter agent manifest on its own, so you can start building an agent without opening the web app first. - Agent memory and browsing access can now be turned on or off by editing and pushing the CLI manifest, instead of only being viewable there. - Agent evals defined in your manifest now actually sync — push, pull, and status all understand the evals block instead of silently ignoring it. ### API - Adding an agent to an orchestration it already belongs to elsewhere now returns a clear error instead of silently creating an inconsistent setup. ## [Claude Fable 5.1 Support](https://brainbaselabs.com/changelog#/r/2026-09-02-claude-fable-5-1-support) 2026-09-02 Claude Fable 5.1 is now available across Brainbase's Anthropic model pickers, featured and ready to use in your agents. ### What's new - Claude Fable 5.1 is featured and marked New across compatible model pickers, placed ahead of Claude Opus 5. - It supports the same adaptive-thinking effort controls (low through max) as other Anthropic models, with a 1M-token context window and up to 128K tokens of output. - Claude Code keeps Claude Opus 5 as its default, so existing agents keep their current behavior unless you switch models. - Usage is billed at $10 per million input tokens and $50 per million output tokens. ## [Muse Spark 1.3 and Gemini 3.8 Flash Support](https://brainbaselabs.com/changelog#/r/2026-09-02-muse-spark-1-3-gemini-3-8-flash-support) 2026-09-02 Meta's Muse Spark 1.3 and Google's Gemini 3.8 Flash are now featured for long-horizon agent and coding work, while their previous generations remain available. ### What's new - Muse Spark 1.3 is now Brainbase's featured Meta model for Kafka and OpenCode agents, retaining native video input and a 1M-token context window while improving agentic and coding performance. - Gemini 3.8 Flash is now Brainbase's featured Google Flash model, with a 1M-token context window, low/medium/high reasoning controls, tool calling, and structured outputs for long-horizon work. - Muse Spark 1.2 and Gemini 3.5 Flash remain selectable under All Models, with no automatic rerouting to the newer versions. ### Pricing - Muse Spark 1.3 usage is billed at $1.25 per million input tokens and $4.25 per million output tokens. - Gemini 3.8 Flash launches at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. - [Meta's Muse Spark 1.3 announcement](https://research.meta.ai/blog/introducing-muse-spark-1-3/) - [Google's Gemini 3.8 Flash announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/) ## [Share Secrets Across an Orchestration's Agents](https://brainbaselabs.com/changelog#/r/2026-09-01-orchestration-secrets) 2026-09-01 Set a secret once on an orchestration and every agent inside it can use it automatically, with per-agent overrides and a clear warning when orchestrations disagree. ### What's new - A new Orchestration → Secrets page lets you set secrets that are automatically available to every agent in that orchestration, instead of adding them to each agent one by one. - Agent secrets pages now show which secrets are inherited from an orchestration, with an Override action to set an agent-specific value instead. - If two orchestrations an agent belongs to set the same secret name to different values, the agent's secrets page flags the conflict instead of silently picking one. - `brainbase agent status` in the CLI now lists inherited secrets and conflicts alongside your local secrets. ### Good to know - Secret values stay write-only everywhere — they're never displayed after saving, and the CLI's local secrets file never receives inherited values. - Removing an agent from an orchestration removes its access to that orchestration's secrets starting with the agent's next turn. ## [GLM 5.3 Model Support](https://brainbaselabs.com/changelog#/r/2026-08-28-glm-5-3-model-support) 2026-08-28 Z.ai's GLM 5.3 flagship is now available in Brainbase's model pickers, with a 1M-token context window and tool calling for long-horizon agentic engineering. ### What's new - GLM 5.3 joins the featured agent model pickers, marked New and promoted alongside the other recent launches. - It supports tool calling and structured outputs, so it works with the same agent tool-use flows as other supported models. Unlike GLM 5.3 Flash, it is a text-only model and does not accept image input. - The context meter recognizes its 1M-token input window and up to 262k tokens of output, so long-context runs are metered accurately. - Usage is billed at Baseten's list rates of $1.40 per million input tokens and $4.40 per million output tokens. ## [Pin an Agent to a Specific Version](https://brainbaselabs.com/changelog#/r/2026-08-28-agent-pinning) 2026-08-28 Agents can now be pinned to a specific saved version, so new tasks keep running on a known-good setup even while you keep editing the agent. ### What's new - Pin any agent to a specific version from its version timeline, the version detail page, or a new Version tile on the agent overview. - While pinned, new tasks run on that version's definition — instructions, entrypoint, installed skills, connected tools, playbooks, and default model — even if the live agent is edited afterward. - Unpin at any time to send new tasks back to the agent's latest version. ### Good to know - A pin freezes the agent's definition, not its operational state: secrets, machine size, memory/browser toggles, and Slack/GitHub/meeting connections always stay current. ## [Agents Can Push to GitHub, Not Just Read It](https://brainbaselabs.com/changelog#/r/2026-08-28-agent-github-write-access) 2026-08-28 GitHub-connected agents can now be granted scoped write access to push branches and commits, so they can open a pull request as part of their work instead of only reading your code. ### What's new - The GitHub integration now has a separate "Git access" capability alongside triggers, off by default for every installation. - When turned on, an agent's sandbox receives short-lived, scoped credentials that can push branches and commits to the connected repository. ### Good to know - Existing installations keep read-only access until you explicitly turn Git access on from the GitHub surface settings. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-08-28-fixes-and-improvements) 2026-08-28 Name and restore agent versions, rename projects, use a redesigned model picker, and see why benchmark cases failed, while managed runs and Slack replies no longer stall or go missing. ### Platform - A managed agent run that stalls mid-task now reliably fails with an alert instead of going quiet, and the original failure reason is preserved end to end. - Right-click delete on an agent or trigger in the workspace scene now asks for confirmation first, and a few related scene and settings bugs are fixed. - Some agent runs previously failed outright if the model provider closed the connection without producing any output; these requests are now automatically retried instead. - The model picker got a full redesign so it's easier to find the right model: browse by Featured, Newly added, or Open-source, or drill into a specific research lab, with per-model pricing shown up front. - A broad visual refresh across the console: redesigned charts, a more interactive agent workspace view, and clean-up across the Usage and Members pages. - Projects can now be renamed from the console instead of being stuck with whatever name they were created with. ### Benchmarks (beta) - Agent Benchmarking (beta) now lets each test case define its own checks and scoring weights, explains exactly why a case passed or failed, and can export a run's full results. - Agent Benchmarking (beta) now walks through run setup in three guided steps, lets you stop a run in progress, and shows exactly why each case passed or failed. ### Integrations - Agent replies triggered from Slack are now tracked until they're delivered, with an automatic fallback message if a run fails to reply, so a request never silently disappears. - A new GitHub integration setting lets your agent ignore other bots' review comments and summaries, so its review isn't anchored to another tool's findings. ### Account and Billing - Displayed pricing for 22 of the 35 models in the catalog now matches what you're actually billed. ### Agents - Any version in an agent's timeline can now be given a memorable name and restored, with accurate attribution for whether a change came from the web app, the CLI, or the API. ## [GLM 5.3 Flash Model Support](https://brainbaselabs.com/changelog#/r/2026-08-26-glm-5-3-flash-model-support) 2026-08-26 Z.ai's GLM 5.3 Flash — released today — is now available in Brainbase's model pickers as a fast, low-cost open-weight option with text, image, and tool-calling support. ### What's new - GLM 5.3 Flash is a Featured model in the agent model pickers, marked New and promoted alongside the other recent launches — available in Brainbase the same day Z.ai released it. - It accepts both text and images and supports tool calling, so it works with the same agent tool-use flows as other supported models. - The context meter recognizes its 1M-token input window and up to 131k tokens of output, so long-context runs are metered accurately. - Usage is billed at Baseten's list rates of $0.15 per million input tokens and $0.50 per million output tokens — roughly a tenth of GLM 5.2's rates. ### About GLM 5.3 Flash - GLM 5.3 Flash is the first natively multimodal model in Z.ai's GLM-5 family: a 320B-parameter mixture-of-experts model that activates 18B parameters per token, with MIT-licensed open weights. Its vision support is built for agent work — reading interfaces, rendered output, spreadsheets, and dashboards. - Before launch it previewed anonymously as "Ox Alpha", the free stealth model that went viral on OpenRouter and OpenCode while the community raced to guess who built it; today Z.ai confirmed they are the same model. - On Z.ai's reported benchmarks it beats GLM 5.2 across the board — 63.4 vs 46.2 on DeepSWE v1.1 and 48.8 vs 26.2 on AutomationBench — and approaches frontier closed models on agentic coding, scoring 84.3 on Terminal Bench 2.1 against Claude Opus 4.8's 85.0. These figures are vendor-reported; independent evaluations are still pending. > 🥷 New stealth model: Ox Alpha > > Ox Alpha is a frontier model built for efficient coding, sustained agentic work, and real-world production use. > > - 1M token context window > - Text, image, and video input > > Try it now and share feedback to improve the model! > > — OpenRouter (@OpenRouter), [on X](https://x.com/OpenRouter/status/2090544970923184269) ## [Members' Access Is Now Scoped to Their Assigned Projects](https://brainbaselabs.com/changelog#/r/2026-08-26-project-scoped-access) 2026-08-26 Standard members now consistently see and act only on agents, workflows, playbooks, skills, and connectors in the projects they've been added to, with the old Viewer role retired in favor of Member. ### What's new - Project-scoped access is now enforced consistently across agents, workflows, playbooks, skills, connectors, and tasks. Organization owners and admins keep full access. - The Viewer role has been retired; existing Viewer members and pending invitations were moved to Member. ### Fixed - Owners and admins can now edit project access for a member who already has at least one project assigned — the edit control had gone missing in that case. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-08-26-developer-updates) 2026-08-26 Manage agent integrations from the CLI, MCP servers configured through the API now reach the agent, and CLI sign-in errors are clearer. ### CLI - CLI commands that fail to authenticate now name which credential actually caused it and give a fix that works for that case, instead of a generic "session expired" or "not signed in." - You can now connect, disconnect, and inspect an agent's Slack and meeting connections, and check MCP server status, directly from the terminal. ### API - MCP servers added to an agent through the API could be saved without error yet never actually be delivered to the running agent; this is now fixed for new and existing agents alike. ## [Agent Benchmarking](https://brainbaselabs.com/changelog#/r/2026-08-25-agent-benchmarking) 2026-08-25 · Beta You can now benchmark an agent's performance directly in Brainbase: build a library of test cases, run them across multiple models, and track score trends over time without leaving the console. ### What's new - A new Benchmarks section on each agent: build a library of test cases, edit them individually or in a spreadsheet-style view, version them, and archive or export a library. - Plan and launch a run across one or more models with configurable repetitions and concurrency, review estimated cost, and confirm before anything starts. - Watch a run execute live, with per-case progress, evidence, and artifacts, and cancel a run in progress. - An overview leads with the agent's most recent score, trend against prior runs, execution coverage, and regressions, with drill-down into individual case results and comparisons against a baseline. - Quality scores and execution health are always shown separately, and partial results always show how many cases they cover. ### Good to know - Contact Brainbase support to have Benchmarking enabled for your team. ## [Project Spend Limits Replace Team Budgets](https://brainbaselabs.com/changelog#/r/2026-08-25-project-spend-limits) 2026-08-25 The Usage page's "Team budgets" are now "Project limits," with a slider to set the cap and a live marker showing exactly where a project stands in the current billing period. ### What's new - Set a project's spend limit with a slider up to the org limit, with a comma-formatted value as you drag. - A marker on the usage chart shows current-period spend against the limit, with a tooltip for details. ## [Agent Version Timeline and Diffs](https://brainbaselabs.com/changelog#/r/2026-08-21-agent-version-timeline) 2026-08-21 Every meaningful edit to an agent is now saved automatically as a version you can browse in a full timeline, with side-by-side diffs showing exactly what changed. ### What's new - Every meaningful edit to an agent — from the console, the CLI, or the API — now creates a version automatically. There's nothing to turn on. - A new Versions page shows the full history for an agent, newest first, including who made each change and when. - Open any version to see a diff: instructions, entrypoint, playbook bodies, function scripts, and mode prompts show line-by-line changes, while other settings show a plain before-and-after. - Each task now shows which agent version it ran on, so you can trace unexpected behavior back to a specific configuration change. ### Good to know - Secret values are never stored in a version — only which secrets exist, never their contents. ## [Generate Images in Agent Tasks](https://brainbaselabs.com/changelog#/r/2026-08-21-agent-image-generation) 2026-08-21 Agents can now generate images directly inside managed tasks using Brainbase-provided models and task-scoped access. ### What's new - Agents can choose from the image models Brainbase exposes, pass supported generation options, and create as many variations as the task requires. - Generated images are saved as ordinary task files and use the existing attachment flow. - The capability is available on newly provisioned managed-agent sandboxes. ### Managed by Brainbase - Brainbase provides short-lived, task-scoped access while keeping provider credentials out of the sandbox. - Image usage is attributed to the task and billed through Brainbase credits. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-08-21-developer-updates) 2026-08-21 List teams and agents and rename tokens from the CLI, sign in more reliably, and push without losing commands, hooks, or files, with token-scope bugs fixed. ### CLI - If an agent manifest declared commands, hooks, or files, `brainbase agent push` could silently leave them out while still reporting success. Push now fails clearly instead of dropping content, and pulling an agent preserves it correctly. - You can now rename an existing CLI token to keep track of what it's used for, and revoking one token no longer signs out a different token stored on your machine. - Creating a token with the `--scope=` or `--scopes=` syntax could silently grant broader access than requested. The CLI now parses both flag forms correctly and rejects malformed scope input outright. - `token create --scope publish` now creates a token scoped exactly as requested instead of a broader one, and installing a skill with a version range now fails clearly instead of silently resolving to nothing. - Signing in with the CLI now uses a short code you approve in the browser, fixing sign-in on Safari and making login work over SSH and inside containers. - Two new read-only commands let you find team and agent IDs from the terminal instead of copying them out of the web app. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-08-21-fixes-and-improvements) 2026-08-21 See what your agent is waiting on, pick follow-up models from curated pickers, and reconnect trigger accounts, while agent runs now survive platform updates. ### Agents - Model pickers now surface the most relevant models first, and you can change the model for a follow-up message right from the composer. - Starting a new task is quicker, and the new-task panel can now be resized to fit how you work. - Agent group pages no longer briefly show agents as standalone before their orchestration membership finishes loading. - Running a SQL query against an agent's memory for the first time could fail outright instead of setting up storage automatically as intended. - On some agent runtimes, a clarifying question from the agent could go unnoticed and the task would incorrectly finish as done; this is now fixed. - Agent turns that are mid-run when Brainbase deploys a platform update now keep going instead of being cut off, and pick back up automatically if their session has to move. - Adding rows to an agent or orchestration memory table that has auto-generated columns, like an identity id, no longer fails with a database error. ### Security - Strengthened default permissions for backend database functions so new ones can't accidentally end up more broadly accessible than intended, and added automated checks to catch this class of misconfiguration going forward. - Fixed a security issue in the page shown after connecting an MCP server that could allow a crafted link to run malicious script in your session. ### Account and Billing - GPT-5.6 Sol usage is discounted 50% now through September 15, 2026 (Pacific time), with the savings shown directly wherever you pick a model, plus a more reliable delivery path for several recent OpenAI models. ### Workflows - You can now disconnect and reconnect the connected account behind an orchestration trigger without losing your trigger setup, and failed trigger deployments now show a clear reason instead of a generic error. ### Integrations - When an agent posts a Slack message on its own initiative, replying in that thread now reliably continues the same task instead of sometimes starting a new one. ### Chat - When an agent pauses to wait on something external, the task list flags it and the task view shows what it's waiting for, with a one-click way to stop waiting. ### Platform - When a Claude model briefly returns a 'model not found' error from its primary path, requests now automatically retry through a backup path instead of failing the turn. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-08-14-fixes-and-improvements) 2026-08-14 Create agent tasks from the CLI, see the current billing period, skip or resume onboarding, and use clearer machine, runtime, and model pickers. ### Agents - Agent setup pickers are more consistent and accurate: machine settings open on click instead of hover, GPU controls only show up where they apply, and more runtimes and models now have their own icons. - Task and thread titles are now generated more consistently and no longer get overwritten or lost once a name has already been set. - Fixed several issues that could cause an agent run to fail outright instead of completing. ### Account and Billing - The Billing page now shows your current billing period and credit refresh date, and the Usage page lets you filter usage stats to that same period. - Usage-based billing and plan entitlements for teams under an enterprise account are now attributed and applied consistently. ### Platform - New users can skip onboarding and finish it later, instead of being forced through it in one sitting to start using Brainbase. - Inviting someone who's already on your team no longer silently fails. ### CLI and API - You can now start an agent task directly from the command line, without opening the dashboard. ## [Meta Muse Spark 1.2 Support](https://brainbaselabs.com/changelog#/r/2026-08-12-muse-spark-native-video-support) 2026-08-12 Meta's Muse Spark 1.2 is now available as a model option for agents, and it's the first model on Brainbase that can natively watch and understand video, right alongside images and documents. ### What's new - Muse Spark 1.2 is available as a model option for agents, with vision support and a large context window. - Muse Spark 1.2 is the first model on Brainbase with native video input — send video straight into a conversation or task, with no separate upload step. - No separate credentials required — Muse Spark routes through Brainbase's centralized model configuration by default. ## [Grok 4.6 Support](https://brainbaselabs.com/changelog#/r/2026-08-12-grok-4-6-support) 2026-08-12 xAI's Grok 4.6 is now the featured Grok model across agents, flows, and chat, with a 500k-token context window, streamed reasoning, and vision support. ### What's new - Grok 4.6 replaces Grok 4.5 as the featured xAI model across agents, flows, and chat; Grok 4.5 remains available under All Models. - The context meter recognizes Grok 4.6's 500k-token input window so long-context runs are counted accurately. - Reasoning effort (low/medium/high), streamed reasoning content, tool calling, image input, and structured outputs are supported end to end. - No xAI credentials required — Grok 4.6 routes through the centralized xAI configuration by default. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-08-07-fixes-and-improvements) 2026-08-07 Memory tables show foreign keys, failed runs explain what went wrong, access controls and secret redaction are stronger, and session replays and long transcripts load reliably. ### Agents - Agents that respond in Slack now know exactly who sent each message, fixing cases where they could confuse a message with someone else's identity — including across different connected Slack workspaces. - When a row write to an agent's memory table is rejected, the memory tab now explains why instead of showing a generic failure. - The memory table editor now shows which column references another table up front, instead of only after a write is rejected. - Canceling a running task now measures usage correctly, and task completion events are delivered in the right order so a run's final status is reported accurately. - The task timeline no longer shows a browser session replay pane, or a misleading playback error, on tasks that never used the browser. - A task whose final step is cut off mid-call is no longer reported as having completed successfully. - New agent onboarding no longer gets stuck bouncing between the creation form and the setup guide when you navigate back or reload mid-setup. - Session recording playback now recovers automatically from a bad video segment instead of freezing, so you don't have to refresh the page to keep watching. - Long task transcripts no longer get stuck on an infinite loading spinner after the task has actually finished successfully. - When a task run or model request fails outright, you now get a clear, specific error explaining what went wrong instead of an unexplained failure. - Agents that use Slack can now reliably open direct messages and read thread replies, fixing two failures that previously blocked those actions. ### Security - We've unified how access to threads, stored files, and attachments is verified across the platform, so it consistently reflects your team's actual membership. - We've broadened automatic secret redaction to provide more thorough, consistent masking of credentials across logs and stored messages. - We tightened how integration credentials and custom secrets are stored and transmitted, so their values are never sent to your browser once saved. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-08-06-developer-updates) 2026-08-06 Tasks created through MCP start automatically, task and thread creation can be retried safely, and skill installs from the registry and GitHub are more reliable. ### CLI - Skills added from the skill registry now install correctly into an agent's sandbox and from the console, instead of silently failing to install. - Skills now install under the name declared in their own instructions, so agents can reliably find and use a skill by the name they were told to use. - Agents now reliably boot with every skill they declare, fixing a case where a skill install could be removed and never replaced. ### API - Fetching an agent, or your full list of agents, no longer fails when one of an agent's skills comes from a GitHub source. - Creating a task or thread through the API is now safe to retry — resending the same request after a network hiccup reuses the original result instead of creating a duplicate. ### MCP - Tasks created through the MCP integration now run automatically by default, matching how tasks behave when created from the app, instead of sitting idle waiting on a start that never happens. ## [Configurable Sandbox Sizing for Agents and Tasks](https://brainbaselabs.com/changelog#/r/2026-08-04-configurable-sandbox-sizing) 2026-08-04 You can now choose how much compute an agent's sandbox uses — from a light default up to a larger tier — and override it for an individual task without changing the agent's setting. ### What's new - Pick a sandbox size (S, M, or L) when creating an agent, each with its own CPU, memory, and disk allocation. - Override the sandbox size for a single task without changing what the agent uses by default. - The agent overview and task views show the active sandbox size, with a quick way to change it after creation. ## [Automatic Model Selection for Agents](https://brainbaselabs.com/changelog#/r/2026-08-03-auto-model-selection) 2026-08-03 Agents can now turn on Auto to let Brainbase pick the best-performing model for each turn automatically, across every supported model or just open-weight ones, instead of you having to choose and manage a model yourself. ### What's new - Turn on Auto (All) to let Brainbase automatically pick the best-performing model for each turn, without exposing which specific model actually ran. - Turn on Auto (Open) to restrict automatic selection to open-weight models only. - Auto is sticky within a task, so once a turn resolves to a model, later turns keep using a consistent, capability-safe choice instead of re-picking at random. - If model resolution can't complete, the run fails visibly instead of silently continuing on an unexpected model. ## [Inline Playback for Browser Agent Sessions](https://brainbaselabs.com/changelog#/r/2026-07-30-inline-session-replay) 2026-07-30 You can now watch a browser agent's recorded session right inside the task view, synced to the event timeline, instead of downloading it first — and full downloads now start instantly with no length cap. ### What's new - A video pane now sits alongside the event timeline, so you can watch what a browser agent did without downloading the recording first. - Click an event to jump the video to that moment, or scrub the video to highlight the matching event — a coverage strip shows which parts of the session have video. - If an event falls outside the recorded coverage, playback lands on the nearest available frame and shows how far off it is. ### What's fixed - Fixed a doubled loading spinner that could appear while a recording was buffering. - Fixed a bug where watching a recording of a still-running task reset the video to the very beginning every 60 seconds. ### Faster full downloads - Downloading the complete recording now starts immediately and streams directly, instead of waiting minutes for the file to be assembled on the server first. - Removed the previous 6-hour cap on how much of a session a single download could cover. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-07-29-fixes-and-improvements) 2026-07-29 Download browser session recordings, and workflow approvals, ChatGPT app connections, and task status are more reliable. ### Agents - You can now download a recording of what your browser agent did during a task, even after its sandbox is gone, so you can review exactly what happened. - Closes a remaining timing gap where a finished task could still show as Running and block you from adding a follow-up. - Browser screenshots in an agent's task history render as images again instead of falling back to raw JSON or a broken-image icon. ### Integrations - Workflow runs that post to Slack no longer include an extra internal status line alongside the real completion message. - Connecting Brainbase to ChatGPT no longer fails because of inconsistent client ID matching, and retrying a connection attempt is now handled safely. ### Workflows - Workflow steps that pause for your approval or additional input now reliably stay paused when an automatic retry occurs, instead of occasionally completing on their own. ## [Kimi K3 Model Support](https://brainbaselabs.com/changelog#/r/2026-07-28-kimi-k3-model-support) 2026-07-28 Moonshot AI's Kimi K3 is now available in Brainbase's model pickers, with support for text, images, and tool calling. ### What's new - Kimi K3 joins Inkling as a new option in the agent model pickers, marked New. - Kimi K3 accepts both text and images and supports tool calling, so it works with the same agent tool-use flows as other supported models. - The context meter recognizes Kimi K3's 1M-token input window and up to 262k tokens of output, so long-context runs are metered accurately. - Kimi K3 is available across Brainbase's cloud coding runtimes, including the OpenCode agent. ## [Shared Memory for Orchestrations](https://brainbaselabs.com/changelog#/r/2026-07-24-orchestration-shared-memory) 2026-07-24 Orchestrations can now give every member agent one shared memory database, edited from a new Memory tab, instead of keeping each agent's memory separate. ### What's new - Turn on shared memory for an orchestration from its settings — every agent that's part of it reads and writes the same memory tables instead of keeping separate private memory. - A new Memory tab on the orchestration gives it the same table-editing experience already available for an individual agent's private memory: browse, query, and edit rows directly. - Shared memory follows the orchestration's existing sharing rules, so only its group members and org admins can turn it on or access it. ## [Claude Opus 5 Support](https://brainbaselabs.com/changelog#/r/2026-07-24-claude-opus-5) 2026-07-24 Claude Opus 5 is now a featured model, with reasoning effort levels through max, a 1M-token context window, and up to 128k output tokens. ### What's new - Claude Opus 5 is listed as a Featured model in the Brainbase model picker for agents, flows, and chat surfaces. - Reasoning effort levels from low through max (including xhigh) are available when Opus 5 is selected. - The context meter recognizes Claude Opus 5's 1M-token input budget and 128k-token max output, so long-context runs are metered accurately. - Claude Opus 4.8 remains available under All Models — existing agents and flows on Opus 4.8 are unchanged. - The new model is supported end-to-end anywhere Brainbase accepts a Claude model selection. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-07-24-fixes-and-improvements) 2026-07-24 Mention playbooks and attach files when starting a task, delete files from the file directory, and send Slack file attachments to coding tasks, with clearer usage totals and task status. ### Agents - Deleting an agent no longer fails with a server error if it has ever received a connector message or had a dedicated data backend provisioned. - The evals panel in a transcript's turn detail view now scrolls properly instead of clipping results once several evals are open. - You can now remove individual files from an agent's shared folder right in the file directory, instead of asking the agent to delete them for you. - Long-running and multi-step agent runs now report a finished status only when the whole run is genuinely done, so a task no longer appears completed while it is still working. ### Chat - Replying to an agent's follow-up question now always resumes the conversation, instead of occasionally leaving the thread stuck showing it still needs info. - You can now @mention a playbook and attach files when creating a task or sending a follow-up, so the agent has the right instructions and materials from the very first message. - While an agent's computer is starting up, the task view now shows a startup notice before the first activity appears, instead of a generic spinner, so it's clear that setup is in progress. ### Account and Billing - The usage page now clearly separates the credits you've spent in the selected time window from your cumulative plan total, so the two numbers no longer look inconsistent at a glance. ### Platform - Enterprise plans once again apply across every organization owned by the same account, so affected workspaces regain their full plan and the ability to invite members. ### Integrations - Files you attach when kicking off a coding task from Slack are now brought into the task workspace, so the agent can work with them from the start. ## [Brainbase MCP: Manage Your Agents from Claude Code and Codex](https://brainbaselabs.com/changelog#/r/2026-07-22-brainbase-mcp) 2026-07-22 Connect Brainbase to your MCP client and manage agents, tasks, evals, schedules, skills, and playbooks directly from tools like Claude Code and Codex, secured with a dedicated OAuth sign-in. ### What's new - Install the Brainbase MCP package into Claude Code, Codex, or the Brainbase registry and control your workspace from your own tools, with no secrets to copy or paste. - A stable set of tools lets you create and update agents, launch and manage tasks, run evaluations, manage schedules, and work with skills and playbooks. - Sign in once with a dedicated, consent-based authorization that only grants MCP access — publishing and administrative actions stay outside its reach. - Safety rails are built in: edits check for the latest version before applying, repeated create calls won't duplicate work, and destructive actions require explicit confirmation. ### Reliability - Filtered search across the agent and skill registries returns matching results reliably. - The Brainbase MCP server now reports an accurate version to connected clients. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-07-18-fixes-and-improvements) 2026-07-18 Browser agents get past CAPTCHAs, teams can share evals and agent settings, agents and tasks can be renamed, and scheduled tasks, Slack threads, and trigger editing are more reliable. ### Agents - Teammates and members of groups an agent is shared with can now view that agent's evaluations and results, instead of seeing an empty evals view. - Editing rows in an agent's memory tables now always updates the row you selected, even when several rows share the same key value. - Recurring schedules set from inside a Claude Code agent now fire reliably in Brainbase cloud sandboxes, so time-based automations actually run instead of being silently accepted and never triggered. - Screenshots and image results from browser and task runs now render as inline thumbnails you can open full-size, and previews that had expired now refresh on their own instead of breaking. - The agent memory viewer now loads table data in a single round trip and lands with aligned headers and row actions, so browsing an agent's memory is quicker and easier to scan. - Any member of an agent's team can now change its model and other preferences — previously only the agent's original creator could, and everyone else hit a permissions error. - GPT-5 and o-series models now run correctly on the OpenCode and Qwen coding runtimes, fixing an error that caused every turn to fail on its first model call. - You can now rename an agent inline from its overview page, with instant validation and no full-page reload. - When an agent stops to ask a question — or a run cuts short before it can finish — the thread and task views surface the question and the reason clearly, with an Answer action right where you're watching the run. - Tasks can be renamed inline from the task detail panel, and the new title updates in place across the task list without waiting for a refetch. ### Integrations - Connecting GitHub now returns you to Brainbase to finish setup, and reconnecting a previously removed GitHub app restores the integration automatically instead of leaving it stuck. - The hosted browser tool now detects and clears CAPTCHA and anti-bot challenges on its own, so browser agents keep moving through sites that used to stop them cold. - Agents connected to Slack now read a thread's history once and then follow only new replies, so long Slack conversations no longer balloon context or cause runs to fail — and duplicate replays are gone. - Slack macro runs now map to a single task across every reply path, and the completion message links directly to that task so you can jump from Slack into the full run in one click. - Editing and testing app triggers is more forgiving: toggle saves no longer wait on slow provider option refreshes, and the Test trigger button gives clear, durable feedback instead of false failures or duplicate submits. ### Workflows - Workflow agents now send their answer as a normal chat message before ending a turn, so replies show up in the conversation instead of turns looking empty. - Running a macro on a workflow that also has webhook or schedule triggers now consistently starts from the macro trigger you chose, and macro invocations fail closed with a clear error if no macro trigger is configured. ### Account and Billing - The free signup credit claim is now reachable from the onboarding wizard and a compact banner in the sidebar, in addition to the Billing page, and disappears everywhere the moment credits are activated. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-07-17-developer-updates) 2026-07-17 Agents created through the API appear in your workspace, the Meeting MCP accepts API keys, unavailable-model errors are clearer, and CLI sign-in is more reliable. ### API - Starting a run or calling the API with a model that isn't available on Brainbase now returns an explicit "model not available" error, instead of a misleading authentication failure or an opaque run crash. - Agents you create through the API now land in a group in your workspace automatically, so they show up in your dashboard alongside the rest of your agents instead of being created out of sight. - The Meeting MCP now accepts Brainbase personal access tokens and live API keys with an explicit profile selection, alongside the existing OAuth flow. ### CLI - The Brainbase CLI now keeps your login fresh across sessions, signs in through the correct Brainbase app domain, and shows the control-plane endpoint you're connected to in login and whoami. ## [Inkling Model Support](https://brainbaselabs.com/changelog#/r/2026-07-15-inkling-model-support) 2026-07-15 Thinking Machines' Inkling is now available for agents and chat — it handles both text and images, and it's free to use for now. ### What's new - Inkling from Thinking Machines is selectable in the agent model pickers, and leads the list for runtimes where it's a strong default. - Inkling accepts both text and images, so you can include screenshots and other visual context in your prompts. - Inkling is free to use for now — turns run at zero credits during this early period. ## [Choose Where Your Cloud Agents Run](https://brainbaselabs.com/changelog#/r/2026-07-15-choose-sandbox-provider) 2026-07-15 You can now pick the sandbox provider for any cloud agent when you create it or change it later, with E2B available as an option for every coding runtime. ### What's new - The sandbox provider picker now appears for every cloud agent you create, so you can choose where the agent's sandbox runs. - E2B is available as a sandbox provider across all coding runtimes. - The picker spells out each provider's trade-offs inline — for example, whether shared folders are available — so you can choose with the details in front of you. - If you don't pick a provider, the existing default is used, and current agents keep the provider they already run on. ## [Choose Your Coding Agent Runtime](https://brainbaselabs.com/changelog#/r/2026-07-15-coding-runtime-picker-ga) 2026-07-15 Pick the coding runtime for any agent — Claude Code, Codex, Cursor, Factory, Qoder, OpenCode, or Qwen — when you create it or change it later. ### What's new - You can choose the runtime both when you create a new agent and when you edit an existing one. - Choose from a single, clearly labeled list of coding runtimes: Claude Code, Codex, Cursor, Factory, Qoder, OpenCode, and Qwen. - Existing agents keep working through the change — their runtime is recognized and shown correctly across your threads, task history, and search. - Bring-your-own-key settings and model options follow the runtime you select. ## [Universal Harness API](https://brainbaselabs.com/changelog#/r/2026-07-14-universal-harness-api) 2026-07-14 One API call takes an inline agent spec and a first message and gives you back a running thread — no separate create-agent step, and identical specs automatically reuse the same agent. ### What's new - Send an inline agent spec — harness, instructions, tools, secrets — alongside your first message and get back a thread id, agent id, and status in a single round trip. - Identical specs automatically reuse the same agent, so reposting the same spec opens a new thread against the existing agent instead of duplicating it. - Prefer an existing agent? Pass an agent id instead of the inline spec — everything else about the endpoint stays the same. - Companion routes for interrupting a run, reading messages, and streaming events live under the same namespace, so you can drive and follow a run without stitching multiple APIs together. - Inline API keys are validated and rejected up front, so a failed key write can never leave behind an agent that reuse would later pick up. ## [A Curated Directory Of Verified MCP Servers](https://brainbaselabs.com/changelog#/r/2026-07-14-mcp-directory) 2026-07-14 Connect agents to verified third-party MCP servers — including Datadog — in a click. The new directory prefills provider metadata, setup guidance, and the exact OAuth callback so most providers are one click away. ### What's new - Agent Tools now has a searchable directory of verified OAuth MCP providers, launching with seven entries including Datadog. - Provider metadata, setup guidance, and the exact OAuth callback are prefilled, so connecting a directory server is a click-through instead of a copy-paste chore. - Custom MCP server setup is unchanged — the directory sits alongside the existing manual flow. - OAuth discovery must succeed before a directory server is added, and transient provider errors surface as retryable failures instead of leaving empty rows behind. ## [Sandbox Runtime Credit Tracking](https://brainbaselabs.com/changelog#/r/2026-07-14-sandbox-runtime-credit-tracking) 2026-07-14 Time spent by managed sandbox tasks now shows up in your credit usage as its own line, with only active minutes counted and team attribution preserved from the run onward. ### What's new - Managed sandbox tasks now consume credits based on active runtime. Only confirmed running time counts — sleeping, starting, archived, and terminal states are free. - Partial minutes stay uncharged while the task can still resume, so brief runs aren't penalized while a task might still come back. - Only sandboxes created after the feature is turned on for your workspace are tracked — existing sandboxes are not back-billed. - Team credit attribution is preserved from the run that used the sandbox, so team billing decisions stay consistent even if usage is reconciled later. - Your Usage dashboard now breaks out sandbox spend as its own total, with daily Chat, Sandbox, and Workflow breakdowns and a per-task view of where sandbox time went. ## [Grok 4.5 Support](https://brainbaselabs.com/changelog#/r/2026-07-13-grok-4-5-support) 2026-07-13 xAI's Grok 4.5 is now selectable across agents, flows, and chat, with a 500k-token context window and streamed reasoning support. ### What's new - Grok 4.5 is available in the model picker for agents, flows, and chat, alongside the `grok-4.5-latest` and `grok-build-latest` aliases. - The context meter recognizes Grok 4.5's 500k-token input window so long-context runs are counted accurately. - Streamed reasoning content from Grok 4.5 is preserved end to end, so tool-using runs get the model's intermediate thinking without extra configuration. - No xAI credentials required — Grok 4.5 routes through the centralized xAI configuration by default. ## [Claude Sonnet 5 Support](https://brainbaselabs.com/changelog#/r/2026-07-10-claude-sonnet-5-support) 2026-07-10 Claude Sonnet 5 is now a featured model, with reasoning effort levels through xhigh, a 1M-token context window, and up to 128k output tokens. ### What's new - Claude Sonnet 5 is listed as a Featured model in the Brainbase model picker for agents, flows, and chat surfaces. - Reasoning effort levels from low through max (including xhigh) are available when Sonnet 5 is selected. - The context meter recognizes Sonnet 5's 1M-token input budget and 128k-token max output, so long-context runs are metered accurately. - Claude Sonnet 4.6 remains available under All Models — existing agents and flows on Sonnet 4.6 are unchanged. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-07-10-fixes-and-improvements) 2026-07-10 Watch agent browser sessions from the task view, run macros from chat and Slack, and get Stripe receipts for auto top-ups. Signup credits now require phone verification. ### Account and Billing - Successful auto-reload top-ups now record the Stripe receipt link so the Billing activity row shows a View receipt button — the same way manual top-ups already do. - Enabling an auto-reload rule immediately checks your prepaid balance and tops up if you're already low, and a background scan catches wallets that dip below their threshold between charges — so top-ups no longer wait for a manual nudge. - New Brainbase accounts activate their free signup credits by verifying a mobile number on the Billing page. Paid top-ups, saved payment methods, and existing balances are unchanged — only the free grant is gated. - Meeting Assistant calls are free through August 15. After the promo, meeting time will bill at 1 credit per started minute — an update from the 10 credits per minute rate announced earlier this week. The promo callout is now visible on the Meetings connector before you enable it. - The old referral promo, invite dialog, and email-based invite flow have been removed. Existing accounts are unaffected — only the legacy invite entry points and their credit granting have been turned off. - When you run out of prepaid credits, the Upgrade plan action in the credits, workflows, and playbooks limit dialogs now opens the in-app Billing page — where you can buy credits, top up, or upgrade — instead of the external payment portal. - Wallet credit top-ups unblock new runs immediately, Team plan upgrades stay active under duplicate payment confirmations, and team credit tracking picks a consistent billing owner across debits and enforcement. ### Agents - When you draw a new connection between two agents, sending a message and continuing an existing task now works out of the box: Send a message is the first tool listed, and messaging tools are enabled by default for new connections. - The developer Task view now has a Browser tab that shows the same per-task browser session already available on live threads, with a view-only default and an explicit Take control mode when you need to interact. - Recurring auto-joined meetings now book exactly one Meeting Assistant bot, so calendars that were edited with "this and following events" no longer end up with two bots joining the same standup. - Switching browser tabs and coming back no longer discards unsaved edits in the agent Instructions editor. Your typed-but-not-yet-saved changes stay exactly as you left them. - Qwen, OpenCode, Cursor, Factory, and Qoder agents now display their official brand icons throughout Brainbase, making it easier to tell your agents apart at a glance in agent lists, thread panels, and side panels. ### Workflows - Macros can now be invoked with a slash command from v2 Chat and the developer Task view, macro-triggered workflows publish cleanly, and Slack macro results are always posted back to the originating thread. - Workflows can be renamed directly from the editor breadcrumb or the /dev topbar with a visible pencil affordance. Renames now persist reliably — including for personal workflows without a team — instead of appearing to save and then reverting. - When you click a run of a managed workflow, it now opens in the agent's task chat surface — where every node's messages, files, and follow-ups live in one thread — instead of a separate legacy view. Older workflow runs still open in their original view. ### Chat - Thread links across search, task lists, kanban, and sidebars now pick the right chat surface based on the thread's runtime — managed cloud and ACP threads open in the agent chat view, while legacy threads keep their existing chat view. - The three-dot actions menu on sidebar threads — including the Delete thread option — is now reachable for every thread, including those near the bottom of a long sidebar list where the menu used to be clipped. ### Integrations - Setting up the Slack integration is now a step-by-step walkthrough: name the app, create it, install it into your workspace, then save credentials — with an inline video demo and a note for workspaces that require admin approval. - Trigger configuration dropdowns that return plain-text option lists — such as the Repository selector when setting up a GitHub trigger — now populate correctly across every trigger and action configuration surface. ### Platform - When an upstream model provider returns a rate-limit or transient error to Brainbase's streaming path, the fallback retry now backs off with jitter and honors Retry-After headers instead of failing the run on the first stumble. ### CLI and API - Organization administrators can pull orchestration manifests through the Brainbase CLI even for groups they are not a direct member of, matching the access rules the rest of the platform already applies. ## [GPT-5.6 Sol, Terra, and Luna](https://brainbaselabs.com/changelog#/r/2026-07-09-gpt-56-model-family) 2026-07-09 Three new GPT-5.6 variants — Sol, Terra, and Luna — are now selectable across agents, flows, and chat surfaces. Sol becomes the default model for new agents and flows. ### What's new - GPT-5.6 Sol, Terra, and Luna are available in the Brainbase model picker for agents, flows, chat, and task surfaces. - Sol is the new default model when you create a new agent or flow. Existing agents and flows keep whatever model you selected before. - Model picker options are now scoped to the runtime family of the surface, so Codex, Cloud, and other runtimes only show models that are compatible with them — no more silently-mismatched selections. - GPT-5.6 tool-using requests are routed with the correct reasoning settings automatically, so tool calls work out of the box. ## [Bring Your Own Key For Cursor, Factory, and Qoder Agents](https://brainbaselabs.com/changelog#/r/2026-07-09-byok-cursor-factory-qoder) 2026-07-09 Cursor, Factory, and Qoder agents now have a purpose-built key-entry flow: enter your vendor API key when you pick one of these harnesses, and Brainbase will prompt you to add or fix the key before a run fails — instead of leaving a mismatched agent to wedge silently. ### Key entry and gating - Picking Cursor, Factory, or Qoder in the new-agent form or the runtime switcher exposes an inline API key field with clear labels for `CURSOR_API_KEY`, `FACTORY_API_KEY`, and `QODER_PERSONAL_ACCESS_TOKEN`, plus a "✓ already set" note when the key is already saved. - Chat and /dev task composers now check the required key before starting a run. If it's missing, a shared key dialog opens, the typed key is saved, and the pending message is resumed automatically once the key is in place. - If the key is removed mid-thread, sending another message opens the same dialog with your draft preserved — you don't lose what you were about to send. ### Better error surfacing - When the runtime rejects a request because a required key is missing, the specific missing key is now surfaced back to the composer instead of a generic failure — no more silent no-op loops for non-owner group members whose secret writes were being blocked. ## [Automatic Wallet Top-Ups When Your Credit Balance Runs Low](https://brainbaselabs.com/changelog#/r/2026-07-08-prepaid-wallet-auto-reload) 2026-07-08 Save a payment method, set a low-balance threshold and a reload amount, and Brainbase will top your prepaid credit wallet up in the background — so agents, workflows, and playbooks keep running without a manual refill. ### What's new - Turn on Auto-Reload from the Billing page once you have a saved payment method. Choose the balance threshold that triggers a top-up and the dollar amount to reload each time. - When your wallet drops below the threshold, Brainbase charges the saved card off-session and adds the credits automatically. Runs in progress are not interrupted while the charge settles. - The Billing page shows a live state for each auto-reload — pending, active, or paused after a failure — so you can see at a glance why a top-up did or did not go through. - If a charge fails, auto-reload pauses briefly before retrying and never double-charges the same top-up; you can also fix the card and resume from the same page. ## [Meeting Assistant Calls Now Bill at 10 Credits Per Minute](https://brainbaselabs.com/changelog#/r/2026-07-07-meeting-assistant-usage-billing) 2026-07-07 Time spent in a Meeting Assistant call is now metered against your credit balance at 10 credits per started minute, so meeting usage shows up alongside your other agent activity in Billing. ### What's new - Every started minute of a Meeting Assistant call is now billed at 10 credits and appears in your usage and credit activity, so you can see meeting time next to the rest of your agent usage. - Only the actual meeting time you attend is billed — pricing is measured from the recorded meeting session, not from calendar invites. - Repeated or delayed usage reports for the same call are settled once and only once, so an in-flight retry can never double-charge a meeting. ## [Prepaid Credit Wallet, Top-Ups, and Team Upgrade Bonus Credits](https://brainbaselabs.com/changelog#/r/2026-07-03-prepaid-credit-wallet) 2026-07-03 Brainbase now runs on a prepaid credit wallet: buy credits directly from a redesigned Billing page, save a payment method for one-click top-ups, and receive a one-time bonus credit grant automatically when you upgrade to a Team plan. ### What's new - A redesigned Billing page shows your credit wallet balance and recent credit activity in one place, with a top-up flow that buys credits in whole-dollar increments. - Saved payment methods make repeat top-ups a single click after your first purchase, without leaving Brainbase. - Team plan upgrades now grant a one-time bonus credit that lands in the wallet automatically once the upgrade completes. - New signup accounts receive a starter credit grant so agents can be tried out immediately. - Enterprise workspaces see a contact path from Billing rather than a self-serve checkout, matching how enterprise contracts are actually negotiated. ### Reliability - Purchases fulfill even if the payment provider's confirmation callback is briefly delayed — the credit lands as soon as you return to Brainbase, and duplicate confirmations are ignored so a top-up never grants twice. - Sandbox-mode and live-mode credits are tracked separately, so switching modes never mixes balances or lets a live purchase count against sandbox usage. ## [Allow, Ask, or Deny Controls for Connected Integrations](https://brainbaselabs.com/changelog#/r/2026-07-03-integration-permission-controls) 2026-07-03 Choose how each connected integration behaves when an agent wants to use it — allow silently, ask for approval, or deny outright — per integration and per action, so agents only pause on the tools that matter. ### What's new - Allow, Ask, and Deny controls now appear on the Tools page next to each connected integration and each of its actions. - Only profile owners and organization admins can change permission rules, so day-to-day teammates can review what's set without accidentally changing it. - The controls stay in sync when you expand or collapse an integration, so the state you see always reflects what will happen the next time an agent tries to use it. ### In-chat - When an agent is waiting on your permission response, the most recent request now stays visible in the conversation even if the agent has moved past that step in its plan, so the approval prompt is easy to find and act on. ## [New Models Available: GLM 5.2, DeepSeek V4 Pro, and Kimi K2.7 Code](https://brainbaselabs.com/changelog#/r/2026-07-03-new-baseten-models) 2026-07-03 Three additional frontier models — GLM 5.2, DeepSeek V4 Pro, and Kimi K2.7 Code — are now selectable in every Brainbase model picker. ### What's new - GLM 5.2 (general reasoning), DeepSeek V4 Pro (general reasoning), and Kimi K2.7 Code (coding) are available in the chat model picker, agent settings, flow nodes, and eval detail views. - GLM 5.2 and DeepSeek V4 Pro reason internally, so no separate effort control is exposed for them — pick the model and use it as-is. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-07-03-fixes-and-improvements) 2026-07-03 Change models mid-chat, send enterprise invoices to a custom billing email, and get more reliable meeting joins and credit tracking. ### Agents - The model picker in the new Brainbase chat now saves your choice on the conversation and applies it on the next turn, so mid-conversation model switches actually take effect instead of being cosmetic. - When an agent joins a meeting and the meeting provider returns a transient rate-limit or capacity response, Brainbase now retries automatically with backoff, so bursts of meeting traffic no longer cause join failures. ### Account and Billing - Enterprise teams can now designate a dedicated billing contact so invoices, receipts, and card-expiry notices go straight to the right accounts-payable inbox instead of the earliest team owner. - Credit usage now captures Codex Cloud and Claude Code Cloud agent activity at the same fidelity as Brainbase cloud agents, and LLM consumption inside sub-threads rolls up into the parent thread reliably. ## [The New Brainbase Experience Is Live](https://brainbaselabs.com/changelog#/r/2026-07-01-new-brainbase-experience) 2026-07-01 Brainbase's redesigned in-product experience is now the default when you sign in, with dedicated Settings, Members, Usage, and Billing surfaces and a refreshed group and agent navigation. The previous interface remains available at /legacy for continuity during the transition. ### What's new - Settings, Members, Usage, and Billing each have their own dedicated pages, replacing the older dialog-based flows so common workspace tasks have a stable URL you can link and bookmark. - Group and agent navigation is unified under a consistent path structure, making it easier to deep-link into a specific agent or share the current view with a teammate. - The Meetings connector settings have been reorganized so the pre- and post-meeting task guidance sits in its own section, and the copy now points directly at the agent's Instructions where that behavior is authored. - Signing in and account bootstrap have been hardened so every workspace ends up with its default Projects group, including for teams that were affected by an earlier race in the account-setup path. ### Continuity - The previous interface is preserved under /legacy for the transition, so links and workflows built around it continue to work while teams move to the new surfaces. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-06-25-fixes-and-improvements) 2026-06-25 Long-running MCP sessions are more resilient, streaming and tool-error recovery are sturdier, and the sign-in pages get a Brainbase refresh. ### Agents - Streamed agent responses recover from upstream failures more gracefully, with automatic retries on early or mid-tool-call failures, clearer error events when something does go wrong, and partial output preserved so future turns stay coherent. - Execution-tool results carry compact failure context for edge cases like non-zero exits and Python tracebacks, helping agents recover from broken commands without confusing read-only logs or file contents for failures. ### Security - The Brainbase naming rollout covers sign-in, browser metadata, MFA completion, and auth emails, keeping the auth path aligned with the rest of the product. ### CLI and API - OAuth-backed MCP connections refresh credentials and reconnect more smoothly during long agent sessions, reducing tool interruptions when provider tokens rotate. ## [Richer File Attachment Handoff in Chat](https://brainbaselabs.com/changelog#/r/2026-06-19-chat-composer-file-attachments-reach-the-agent) 2026-06-19 Chat composer uploads include workspace paths, startup retries, and first-message handoff support, making dragged, pasted, and selected files more reliable context for agent runs. ### What's improved - Files added to a chat message are uploaded into the task workspace and referenced with their workspace paths, giving the agent direct context for file and vision tools. - Uploads wait through sandbox startup when a brand-new chat is still getting ready, while the composer keeps a visible starting state. - If an upload fails, the composer preserves the text and selected files so the user can retry without posting a message with missing context. - Attachments on the first message of a new thread are carried through the navigation into that thread before sending. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-06-19-fixes-and-improvements) 2026-06-19 Edit sandbox entrypoints and browse each agent's filesystem in the developer app, chat gets date headers, a thinking indicator, and bookmarks, and long interactive sessions cost less to cache. ### Agents - New accounts landing on the developer app now go through a short multi-step welcome wizard, and the Help → Tutorials experience matches it with the same full-page, decluttered Web and Code walkthroughs. - Each agent's persistent file system now has its own Filesystem tab in the developer sidebar — next to Memory and Browser — with a polished native file browser instead of an inline section on the overview page. - New accounts landing on the developer app now see a guided UI-or-CLI onboarding chooser, and every developer page has a persistent help button linking to docs, an in-app tutorial wizard, and contact. - Orchestration connections can now expose multiple dispatch tools to the source agent and run in both directions, with the two directions drawn as parallel lanes on the canvas. - The bash script that prepares an agent's sandbox before the harness boots is now editable from a new Entrypoint tab in the developer sidebar, alongside Secrets and Instructions. - The tasks list and overview on an agent's developer page now show all of that agent's threads instead of a small slice, even when the agent's team is busy. - Prompt caching now survives multi-minute idle gaps in human-interactive threads, and team-connected agents stop paying a full prefix rewrite on every new turn. - Scraped tool or web content that contained an embedded null character no longer aborts the surrounding agent run with a database write error. ### Chat - Permission-only pauses in chat are more direct: the approval card appears without an extra free-text step, while multi-question need-more-info flows still behave the same way. - The Files tab on the agent chat now loads the same way it does in the developer task view, so chats with cloud-runtime tasks no longer sit on a loading state forever. - Several long-asked polish items now ship on the agent chat: per-day date headers, a thinking indicator during sandbox spin-up, a bookmarks panel, owner attribution for teammates, and a queued send-while-busy flow so mid-turn sends no longer drop or show the wrong banner. ## [Enterprise Billing: Invoices, Saved Cards, and Live Payment Status](https://brainbaselabs.com/changelog#/r/2026-06-18-enterprise-billing-invoices-and-payment-methods) 2026-06-18 Enterprise customers can now view their invoices, add a card on file, and see live payment status without leaving Brainbase. ### What's new - The billing tab now shows a full invoices list with per-line classification into Plan charges, Overages, and Other, each linking out to the hosted invoice. - Enterprise admins can add or replace the card on file directly from the contract card via a hosted checkout flow, with on-file status surfaced inline. - Invoice and subscription status update in real time as payments succeed, fail, or are voided, so the billing tab reflects the same state your finance team sees. - Contracts that bill due-on-receipt now require a card on file before they can be promoted to fully migrated, preventing silent payment gaps at cutover. ## [Agent as a Judge](https://brainbaselabs.com/changelog#/r/2026-06-18-agent-as-a-judge) 2026-06-18 Evals can now be judged by a full agent run instead of a single-shot model — the judge can open files, browse artifacts, and reason the same way the agent under test does. ### What's new - Every eval row in the developer app has a new Model / Agent toggle and an agent picker, so each eval can choose between a quick single-shot model judge and a full agent judge per rubric. - Agent judges run as their own task, which means they can open files in the workspace, inspect artifacts, and follow multi-step reasoning — the kinds of checks a one-shot model judge can't reliably do. - Verdict cards in both the turn view and the eval detail modal now show which agent rendered the verdict, with a View judge run link that opens a read-only transcript of the judge's reasoning. - From the judge-run modal, Open full view expands the transcript inline into the read-only task detail, so reviewers can dig into the judge's tool calls and intermediate steps without leaving the page. - Judge runs are filtered out of the regular task list so evaluation traffic doesn't pollute your conversation history. ## [The File System](https://brainbaselabs.com/changelog#/r/2026-06-18-the-file-system) 2026-06-18 Each agent now has its own persistent file system — a shared workspace that survives across every thread and every run for that agent, so files written by one conversation are there for the next one to read. ### What's new - Every agent now has a persistent file system that lives across all of its threads and runs. Anything the agent writes to its workspace in one conversation is available the next time it (or you) opens a new thread for the same agent. - A per-agent toggle on the developer agent overview turns persistent storage on or off; the same control is available on the enterprise agent detail view, scoped to the selected agent. - File browsing, uploads, and downloads on cloud-runtime agents all go through the persistent workspace, so the developer file tree and the in-thread Files tab show the same shared state. - Toggling the setting is atomic, so flipping it on multiple agents in quick succession no longer races and reverts other recent preference changes. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-06-18-developer-updates) 2026-06-18 Name and rename API keys, see MCP server health in the developer app, and pull agents without duplicated marketplace skills. ### API - API keys can now be given a human-readable name when created and renamed inline later, so it's easier to tell production, staging, and per-developer keys apart. ### MCP - The developer Tools page and per-run task detail now show live health for each MCP server, with separate states for connected, not yet checked, and "couldn't verify this run" so an unverified or unhealthy server is no longer painted as healthy. ### CLI - Installing a marketplace skill from the UI no longer surfaces twice in the CLI manifest, so `agent pull` and `agent push` work cleanly without an unusable placeholder entry. ## [Mobile-Friendly Agent Chat and Developer Surfaces](https://brainbaselabs.com/changelog#/r/2026-06-16-mobile-experience-for-dev-and-chat) 2026-06-16 The agent chat and developer surfaces now have proper small-screen layouts: a slide-out drawer for navigation, a full-width conversation, and a composer pinned to the bottom of the screen. ### What's new - The agent chat surface now reflows on phones: the task-list sidebar becomes an off-canvas drawer, the conversation goes full-width, and the side execution panel is hidden where it cannot fit. - The developer surface gains a hamburger-opened drawer with a scrim and a persistent top bar, so every page is reachable from a phone. - Task and chat detail views go full-width with the composer pinned to the bottom and safe-area aware; header actions collapse to icons so rows do not overflow. - List views reflow from dense tables into stacked cards on small screens, and oversized grids scroll horizontally instead of clipping. - Input fields use a font size that prevents iOS from auto-zooming on focus, and the layout uses the dynamic viewport height so the composer is not hidden behind browser chrome. ## [Gemini 3.5 Flash Support](https://brainbaselabs.com/changelog#/r/2026-06-12-gemini-3-5-flash-support) 2026-06-12 Gemini 3.5 Flash is now available as the new free-tier Google model, replacing discontinued Gemini options that Google no longer offers. ### What's new - Brainbase now supports Gemini 3.5 Flash as the free-tier Google model for agents. - Removed Gemini models that Google has discontinued: Gemini 3 Pro Preview, Gemini 3.1 Flash Lite Preview, and Gemini 2.0 Flash. - Other Google models stay available: Gemini 3 Flash Preview, Gemini 2.5 Pro, and Gemini 2.5 Flash. - Workflow schema documentation and examples now reference Gemini 3.5 Flash as the default Google example. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-06-12-developer-updates) 2026-06-12 Push playbooks with `brainbase agent push`, unpinned registry skills stay up to date, and the Meeting MCP installs for connected agents. ### CLI - `brainbase agent push` now writes playbook components alongside instructions, skills, and MCPs, so playbooks can live in source control and reconcile to your cloud agent in a single push. - When an agent references a registry skill without pinning a version, each Brainbase CLI push now resolves to the latest published version. ### MCP - Agents whose workspace has an active meeting integration now automatically get the built-in Meeting MCP installed, matching how Slack-connected agents already work. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-06-12-fixes-and-improvements) 2026-06-12 Meeting bots use your saved name and avatar, Stop now cancels the running workflow, and team settings open from group pages. ### Agents - Tools that need explicit user approval now consistently surface an approval prompt with the correct capability attached, instead of relying on the model to recover from the error mid-turn. - The bot name and avatar saved in Meeting Setup now carry through to the meeting attendee, so participants see the identity you configured. ### Platform - Organization-level settings tabs now appear in the settings dialog from any page that shows the agents sidebar, not just the team home, so admins can adjust team settings without backing out first. ### Workflows - Pressing Stop in a workflow-driven thread now cancels the workflow itself, so the run cannot resume later and finish in the background. ### Integrations - Slack threads no longer receive a redundant completion summary after a workflow finishes. The agent's own posted result remains the final message. ## [Claude Code and Codex Usage Now Counts Toward Credits](https://brainbaselabs.com/changelog#/r/2026-06-11-subagent-credit-tracking) 2026-06-11 Claude Code and Codex subagent runs now appear in team credits and usage reporting, bringing coding-agent spend into the same view as the rest of your model activity. ### What's new - Claude Code and Codex subagent turns are recorded against the parent thread's billing owner with the same token-based pricing as other model calls. - Usage dashboards and usage-billed plans now include coding-agent spend instead of leaving those runs unmetered. - Running threads are unaffected. This update changes reporting and billing attribution, not how the coding agents execute. - Team credit-limit checks before a subagent starts will follow separately; this release covers tracking after the run completes. ## [Faster Long Conversations in the Thread View](https://brainbaselabs.com/changelog#/r/2026-06-10-faster-long-conversations) 2026-06-10 Long threads stay smoother: idle background checks no longer redraw the whole conversation, and streaming only updates the message currently changing. ### What's better - Background checks for bookmarks, token usage, and entitlements now leave the thread alone when nothing changed. - Streaming responses update the active message and running tool output instead of repainting every message on every chunk. - Auth and profile refreshes, including hourly token refreshes, no longer cascade through every message in the open thread. - Very large threads can still be limited by the size of the rendered conversation itself; deeper virtualization is a separate follow-up. ## [Claude Fable 5 Support](https://brainbaselabs.com/changelog#/r/2026-06-10-claude-fable-5-support) 2026-06-10 Claude Fable 5 is now available, with concise reasoning summaries enabled by default and four effort levels to choose from. ### What's new - Brainbase now supports Claude Fable 5 for agents that use Anthropic models. - Adaptive thinking now surfaces concise summaries by default, so agents can explain their work without dumping internal reasoning. - Low, medium, high, and max effort levels are supported, letting each agent trade speed for deeper reasoning as needed. - Existing agents continue on their previously selected model; switching to Fable 5 is an explicit per-agent change. ## [Coding Agents Resume Slack and Teams Tasks](https://brainbaselabs.com/changelog#/r/2026-06-05-slack-coding-agent-task-continuation) 2026-06-05 Follow-up messages to Claude Code or Codex in Slack and Teams now continue the same coding task instead of starting a new one every time. ### What's new - The first message in a Slack or Teams thread creates the coding task; later messages in that thread append to and resume the same task. - Two follow-up messages that arrive at the same time no longer create duplicate tasks. - Coding agents no longer create an extra orphan thread, so the Tasks view stops showing phantom running rows for each coding conversation. - Standard chat agents and Slack workflow commands are unchanged. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-06-05-fixes-and-improvements) 2026-06-05 Turn memory on or off per agent, manage browser proxies, and fill in workflow inputs inline in Slack, with faster app triggers and no more duplicate or skipped workflow runs. ### Workflows - Starting the same workflow again from Slack or the dashboard now creates a fresh run, even when the inputs match a recent run. - Workflow agent steps now recover from transient timeouts and stay focused on the current step before handing control back to the workflow. - Gmail, Linear, and other app triggers now collapse duplicate deliveries into one workflow run, so one source event does not create multiple threads. - App triggers that rely on polling now check for new events about every 15 seconds instead of every 15 minutes. - Threads started from Slack now use the workflow's naming convention instead of staying named after the slash command. - Slack-triggered workflows now finish with the workflow's final output instead of re-posting an earlier acknowledgement. ### Agents - Unreadable or truncated images now produce a clear error after one retry, instead of failing the entire agent run. - Persistent memory can now be enabled or disabled per agent from the sidebar, with a confirmation step before memory is turned off. - Configure HTTP and HTTPS proxies for an agent's browser sessions, including regional egress, authenticated networks, and staging environments. - The orchestration view remembers your camera angle, and agent pages opened from an orchestration keep a breadcrumb back to that orchestration. ### Integrations - Slack workflow commands can now collect inputs directly in the thread with a Run button, alongside the existing modal flow. - Integration setup forms now hide internal fields, so triggers and actions only ask for the information your workflow actually needs. - When you start a workflow from Slack, the confirmation message now includes the inputs you submitted so you can verify the run without leaving Slack. ## [Start Coding Agents From Slack and Teams](https://brainbaselabs.com/changelog#/r/2026-06-04-coding-runtime-agent-dispatch) 2026-06-04 Slack and Teams messages can now start Claude Code and Codex agents as background coding tasks that post results back to the thread. ### What's new - When an agent uses Claude Code, Codex, or a Brainbase cloud-coding runtime, Slack and Teams mentions dispatch a background task instead of opening a synchronous chat stream. - The task keeps the original thread context, so results return to the same Slack or Teams conversation. - Long-running coding sessions no longer tie up the chat handler while they work. - Standard chat agents are unchanged and continue to stream replies inline. ## [Developer Updates](https://brainbaselabs.com/changelog#/r/2026-06-04-developer-updates) 2026-06-04 See orchestration triggers, attach your published skills, and sync UI-built orchestrations from the CLI. ### CLI - The Brainbase CLI can now show orchestration triggers, including app triggers such as Linear issue updates. - Skills you publish with the Brainbase CLI can now be attached to agents and appear in the UI, so published custom skills are ready to use. - Orchestrations created in the UI now pull and push cleanly through the Brainbase CLI, including members that did not start from a local checkout. ## [Invite Kafka to Calendar Meetings](https://brainbaselabs.com/changelog#/r/2026-06-03-meeting-calendar-auto-join) 2026-06-03 Give an agent a calendar email, invite it to a meeting, and it will join on time, capture the conversation, and bring the transcript back into the thread for summaries, follow-ups, and next steps. ### What's new - Generate a dedicated calendar email from the meeting setup dialog and CC or invite it to any Google or Outlook calendar event. - Set a lead time of at least 10 minutes so Kafka has enough notice to join at the scheduled start. - After the call, the transcript is added to the resumed thread so the agent can summarize, file follow-ups, or run workflows using what was actually said. - Calendar auto-join is opt-in per agent. Agents without a calendar email configured behave exactly as before. ## [Custom Modes](https://brainbaselabs.com/changelog#/r/2026-05-29-custom-agent-modes) 2026-05-29 Create custom modes with your own names and transitions, backed by stricter validation and clearer instructions for the model. ### What's new - Custom mode names and transitions are validated before they are used. - Mode summaries sent to the model are normalized for consistency across runs. - Setup and onboarding flows skip unnecessary custom-mode lookups so those paths stay lightweight. ## [Fixes and Improvements](https://brainbaselabs.com/changelog#/r/2026-05-29-fixes-and-improvements) 2026-05-29 Reassign agent ownership, keep model preferences per thread, preview past workflow versions, and see enterprise contract details in Billing, with workflows that pause, resume, and cancel cleanly. ### Workflows - Workflows started from Slack or the web app now resume correctly after a reviewer approves a paused step. - Kafka workflows now wait for missing input instead of failing, then resume when the needed details arrive. - Version history now shows who made each meaningful edit across agents, playbooks, and workflows. - A Google-Docs-style preview for past Kafka workflow versions: click a version in history, view it on the canvas hermetically, then Restore or Exit. ### Agents - Turn on recovery mode to block risky follow-up actions after a tool fails, while keeping read-only inspection available so the agent can understand what happened first. - SVG and other unusual image attachments now convert or fall back gracefully instead of breaking the message. - Team admins and owners can transfer an agent to another team member directly from the agent side panel. ### Account and Billing - Enterprise admins and owners can review contract terms directly in Billing without asking Brainbase staff. ### Integrations - Kafka now recovers from desynced external trigger deployments and gives clearer errors when publishing cannot proceed. ### Chat - The model picker remembers the chat model you chose for each thread, so opening it later picks up where you left off instead of resetting to the agent default. ## [Claude Opus 4.8 Support in Kafka](https://brainbaselabs.com/changelog#/r/2026-05-28-claude-opus-4-8) 2026-05-28 Kafka now supports Anthropic's newest Opus model, including low-to-max effort controls and Fast mode. ### What's new - Claude Opus 4.8 is available as a featured Kafka model with low / medium / high / xhigh / max effort levels and Fast mode. - Claude Opus 4.7 remains available under the other Anthropic models section so existing agents keep working unchanged. - The new model is supported end-to-end anywhere Kafka accepts a Claude model selection. ## [Slack Workflow Commands](https://brainbaselabs.com/changelog#/r/2026-05-28-slack-workflow-macros) 2026-05-28 Teams can launch Kafka workflows directly from Slack with custom slash commands and reliable one-response delivery. ### What's new - Workflow trigger nodes now accept a Slack Command — the slash command users type in Slack to start the workflow. - Kafka-managed Slack apps include the interactivity needed for macro input buttons and modals. - Macro deliveries use durable delivery so each run replies exactly once, even on transient Slack failures. - Modals work for legacy Slack installs without reinstalling the app. ## [Enterprise SSO for Kafka Teams](https://brainbaselabs.com/changelog#/r/2026-05-27-customer-sso-saml) 2026-05-27 Enterprise teams can connect their own SAML identity provider, auto-provision users on first sign-in, and merge matching existing accounts. ### What's new - Sign in with SAML through the identity provider your team already manages. - First sign-in JIT-provisions the user into the team that owns the SAML connection. - Existing non-SSO accounts on the same email can be merged into the new SSO identity with an audit trail. - Sign-in UX detects email domains, surfaces a Continue with SSO button, and blocks password sign-in for SSO-enforced domains. - Operations teams get safer onboarding, removal, rollback, and drift-check flows for SSO domains.