What's New

Self-serve account deletion with a 14-day grace period

Users can now delete their accounts with a 14-day grace period to reconsider. Desktop sessions can switch to an Orchestrate view for managing agents. A new page allows BedrockNews players to claim credits after playing. Various fixes improve the user experience, including message retention, document scrolling, and chat settings application.

• New: Organizations can now delete their accounts from Settings with a 14-day grace period to cancel the deletion.
• New: Desktop sessions that start other agents can now switch to an Orchestrate view to manage all agents at once.
• New: A new page allows BedrockNews players to claim their credits after playing.
• Fixed: A message you start typing now stays with the conversation you typed it in, even when switching conversations.
• Fixed: The desktop window no longer scrolls into empty space, improving navigation and visibility.
• Fixed: Chat settings now apply only to the specific chat they were changed in, enhancing user control.
• Fixed: Importing a ChatGPT or Claude history file that is invalid will now be marked as failed, preventing blocks.
• Fixed: Users will now receive a warning if Ollama in WSL limits GPU memory, improving awareness.

New trusted-account tier increases spend ceiling for organizations

Organizations can now benefit from a higher spend ceiling with the new trusted-account tier. Additionally, organizations can review payment-risk alerts as detections.

• New: Organizations can now benefit from a higher spend ceiling with the new trusted-account tier.
• New: Organizations can now review payment-risk alerts as detections in the admin panel.

Message any session across projects with new desktop feature

Desktop conversations can now message any other conversation across projects, with approval holds and @-mentions. Organizations can also use capped grant packages for invite signups and see referral rewards. Other updates include TTS voice samples in Studio, improved billing notifications, and better error messages for file attachments.

• New: Desktop conversations can now message any other conversation across projects, with approval holds and @-mentions.
• New: Organizations can now use capped grant packages for invite signups.
• New: Inviters can now see their referral rewards and receive notifications about them.
• New: The Speech tab in Studio now lists every TTS voice with a play button for samples.
• Fixed: When a file attachment in chat is refused, you now see a notification explaining why.
• Fixed: Fetching a link to an unreadable file type now explains the issue instead of failing with a server error.
• Fixed: When your credit balance can't cover a full answer, it is now shortened with a note instead of going negative.
• New: Buying a credit pack now releases the remaining signup grant schedule.
• New: Organizations can now perform per-user signup grant admin actions.
• New: Image uploads are now moderated based on a per-user daily allowance of detector calls.
• Fixed: The new-subscriber Slack alert now shows subscriber names as plain text.
• Fixed: Cancelling a GitHub connection now shows a specific cancellation message instead of a generic error.

Sign in with Apple now available for organizations

Organizations can now enable Sign in with Apple for user authentication. Additionally, users can attach PDF files to chats and add folders from the composer. Alerts for credit spend spikes and a daily credit drip for new accounts have been introduced. The chat model picker has been improved with new features, including showing cached and pending model server lists. Other enhancements include better handling of tool failures in chats and improved analytics for purchases and internal accounts.

• New: Organizations can now enable Sign in with Apple for user authentication.
• New: Users can now attach PDF files to chats in Bike4Mind Desktop.
• New: Users can now add folders to a chat from the composer menu.
• New: Organizations will receive alerts on credit spend spikes.
• New: New accounts will now receive a daily credit drip based on a schedule.
• Improved: The chat model picker will now show cached and pending model server lists while refreshing.
• Fixed: When a tool fails in a chat with an OpenAI model, the assistant now explains the failure.
• New: Purchases will now be associated with the visit in which they were made.
• New: Internal accounts' events will now be marked to exclude them from customer counts.
• New: Subscription purchase events will now include the end date of the subscription period.

Download the new Windows Desktop installer for all architectures

Users can now download a Windows installer for the latest desktop release, which supports both x64 and arm64 architectures. This simplifies the installation process for Windows users.

• New: The Desktop download page now offers a Windows installer for the latest desktop release.
• New: Desktop releases now include a single Windows installer that supports both x64 and arm64 builds.

Chat with model servers over SSH on Desktop

The Desktop app now allows users to chat with model servers over SSH. It also provides installation guidance for Ollama and displays model pull locations. Other improvements include clearer messaging for tool failures and enhanced billing metadata for risk assessment.

• New: Desktop can now add a model server by its SSH address and chat with its Ollama through an SSH client.
• New: When Ollama does not answer on this computer, the Models screen now informs the user and offers steps to resolve the issue.
• New: The pull dialog now shows where a model will be saved and how much space is free.
• New: Installed models on the Models screen now have a 'Use in chat' button to start a new conversation.
• Fixed: When a fetched web page has no readable text, the assistant now suggests another source instead of reporting a tool failure.
• Fixed: When a request uses a model your plan does not include, the assistant now informs you immediately.
• New: Stripe now receives account age and signup network as metadata for risk assessment.

Unlink expired connections directly from integrations

Organizations can now unlink expired Atlassian or Notion connections directly from the integrations section. CSV files in the file browser now display a table icon instead of a generic JSON icon. The Send button in chat now has an accessible name for better screen reader support.

• Fixed: Organizations can now unlink expired Atlassian or Notion connections directly from the integrations section.
• Fixed: CSV files in the file browser now display a table icon instead of a generic JSON icon.
• Fixed: The Send button in chat now has an accessible name for better screen reader support.

Post messages to Slack channels with new API key feature

Organizations can now enable Slack channels for API posting, allowing messages and replies to be sent as the Bike4Mind bot. The CLI has been improved with a preselected hosted server on fresh installs, a pre-filled sign-in link, and input history retention across sessions. Additionally, alarm notifications now include categories, and cost alarms can be sent to a dedicated Slack channel.

• New: Organizations can now enable Slack channels for API posting, allowing messages and replies to be sent as the Bike4Mind bot.
• Improved: On a fresh install, the CLI now preselects the hosted server for easier sign-in.
• Improved: The CLI sign-in screen now provides a pre-filled link for easier access.
• Improved: The CLI now remembers input history across sessions, making it easier to reuse commands.
• Improved: The CLI can now access and manage model servers saved in the Desktop application.
• New: Alarm notifications now include categories, and cost alarms can be sent to a dedicated Slack channel.

New accounts must verify email before sending invites or using tools

New accounts are now required to verify their email and meet minimum age requirements before accessing certain features. The desktop app has been restyled, and the left navigation now organizes items more efficiently. New alarms for monitoring young-account LLM spending and credit overdrafts have been added. Additionally, usage events will now track the funding source for credits spent, enhancing transparency.

• New: New accounts must verify their email and reach the minimum account age before using key features.
• Improved: The 'Try the desktop app' suggestion in the sidebar has a new clickable design.
• Improved: The left navigation now keeps up to eight items in view, moving the rest into a More menu.
• New: New alarms have been added to monitor young-account LLM spending and credit overdrafts.
• New: Usage events will now record the funding source for credits spent, improving tracking.

Require 3-D Secure for first purchases on new accounts

New accounts will now require 3-D Secure confirmation for their first purchase. Additionally, organizations will no longer receive incorrect alerts for credit purchases when the owner has been deleted.

• New: The first purchase on a new account may ask you to confirm the payment with your bank (3-D Secure).
• Fixed: Organizations will no longer see incorrect alerts for credit purchases when the owner has been deleted.

Show session context breakdown in the session panel

The session panel now displays how context is utilized, including tools and project files. Organizations can enable a feature for rating answers to earn credits. Free credits now include descriptions of their sources, and agents using older Anthropic models will now run on Claude Opus 4.8 instead of failing. Completion rounds are capped at the remaining balance to prevent overdrawing credits. The sidebar now highlights the 'Start here' tag in blue for better visibility.

• New: The session panel now shows how context is used, including tools and project files.
• New: Organizations can now enable a feature that allows verified users to rate two model answers and earn promotional credits.
• Improved: Free credits in your credit history now say where they came from instead of 'Generic credit addition.'
• Fixed: Agents set to Claude Opus 4 or Opus 4.1 now run on Claude Opus 4.8 instead of failing.
• Fixed: When your credit balance is nearly used up, replies may be shorter rather than taking your balance below zero.
• New: The 'Start here' tag on the Gears row in the sidebar is now blue for better visibility.

Pay failed renewal invoices directly from the billing page

Users can now pay failed renewal invoices immediately from the billing page. Additionally, desktop sessions can relay messages more efficiently, and unsupported file formats will now log warnings instead of errors. Other improvements include better handling of invite request limits and subscriber email storage.

• New: Users can now pay failed renewal invoices immediately from the billing page.
• New: A message sent from one desktop session to another is now delivered after the current step.
• Fixed: Unsupported file formats now log warnings instead of unnecessary error alerts.
• Fixed: Invite request limits are now enforced for all invite routes.
• Fixed: Subscriber emails are now stored in a normalized format to prevent duplicates.

Edit user profiles directly from the user profile page

Admins can now edit a user's name, username, and email directly from their profile, with email changes requiring confirmation. The model picker for free accounts now shows paid-only models as locked. Emailing files is limited to 5 recipients per email, with hourly and daily sending limits. The Filters panel in the sidebar now opens centered under its button and indicates when more content is scrollable. Access to Invites has been added to the People sidebar under the More section.

• New: Admins can edit a user's name, username, and email from the user profile, with email changes requiring confirmation.
• Improved: The model picker for free accounts now lists paid-only models as locked, with an Upgrade button.
• Fixed: Emailing files or notebooks now allows up to 5 recipients per email, with sending limits per hour and per day.
• Fixed: The Filters panel in the sidebar now opens centered under its button and shows when there is more to scroll.
• Fixed: The People sidebar now includes access to Invites under the More section.

API reference now documents promptText for chat replies

The API reference has been updated to clarify that `promptText` is returned as an object in chat replies with `wait: true`. Additionally, the What's New headline and summary will be trimmed if they exceed the character limit. Cache reads for the Claude model will now be billed at their published rates.

• Fixed: The API reference now documents `promptText` on a `wait: true` chat reply as an object, not a string.
• Fixed: The What's New headline and summary will now be trimmed if they are too long.
• Fixed: Cache reads for the Claude model will now be billed at their published rates.

New chat creation suggestions in the desktop app

The desktop app now opens new chats with example prompts for various media types. Organizations can also rate curated picks based on hardware settings. Tooltips have improved readability, and signing in defaults to the production server.

• Fixed: Signing in to the desktop app now starts on the Bike4Mind server with nothing to type.
• New: New chats in the desktop app now open with example prompts for images, video, voice, sound effects, music and writing.
• New: Organizations can now rate curated picks and fit badges based on hardware specifications entered in the settings.
• Improved: Desktop tooltips use an opaque background so their text stays readable over other content.
• New: The Scopes reference in API key settings now lists every endpoint each scope unlocks, grouped by area.

Share one CloudFront viewer-response function across preview stages

Organizations can now share a CloudFront viewer-response function across preview stages, reducing the number of functions needed.

• New: Organizations can now share one CloudFront viewer-response function across preview stages, reducing the number of functions needed.

New daily budget for free signup grants

New accounts unlock their free credits gradually based on daily usage. Users can now see previews of pasted and attached files in chat. Buying credits allows access to paid models and personal API keys. Accounts awaiting signup review temporarily lose AI access. Long answers on thinking models no longer get cut off. The embed feature now rejects opaque origins for improved security.

• New: New accounts unlock the rest of their free credits gradually based on daily usage.
• Improved: Pasted and attached files now appear as previews in the message box with upload progress.
• Fixed: Buying credits now unlocks access to paid models and personal API keys.
• Fixed: Accounts awaiting signup review temporarily lose AI access while keeping existing data access.
• Fixed: Long answers on thinking models no longer get cut off and restarted after six minutes.
• Fixed: The embed feature now rejects opaque origins on every embed key for improved security.

Free-tier accounts now restricted to allowed models

Organizations can now ensure that free-tier accounts only have access to allowed models. This change prevents the Default API Model from bypassing the free-tier model gate.

• Fixed: Organizations can now ensure that free-tier accounts only have access to allowed models.

Free accounts can now access specific media models

Free accounts can use the media models explicitly included in the free-tier model list. The sign-in process will now refuse requests with a misconfigured Turnstile gate, improving security during sign-in.

• Fixed: Free accounts can use the media models explicitly included in the free-tier model list.
• Fixed: The sign-in process will now refuse requests with a misconfigured Turnstile gate, improving security during sign-in.

See model fit ratings for your machine in the CLI

Organizations can now view a FIT rating for each model build in the CLI, indicating suitability for the machine. Additionally, signing up with Google or GitHub now requires a verified email address. Stranded file references are dropped after project or session revokes, eliminating unnecessary warnings.

• New: Organizations can now see a FIT rating for each model build, indicating whether it is suitable for the machine.
• Fixed: Signing up with Google or GitHub now needs a verified email address from that provider.
• Fixed: When you leave a project or a project is unshared, files you can no longer read are removed from your pinned files.

New model filtering and tailored recommendations on the desktop

The desktop now features a Curated models tab with tailored recommendations and a new model picker that lists the newest models first. Additionally, the Credits ledger can be filtered by model, source, and date. Other improvements include a clearer sign-in verification dialog, separate MFA trusted device memory for each account, draggable AI chat pill, and better handling of free-tier daily limits.

• New: The Models screen now features a Curated tab with tailored model recommendations and fit badges for your computer's memory.
• New: The Credits ledger can be filtered by model, source, and date, with a spend breakdown by model or source.
• Improved: The desktop model picker now lists the newest models first and no longer shows each model's internal id.
• Fixed: The "verify you are human" check on sign-in now shows in a clear dialog with a short explanation and a Cancel button.
• Fixed: If you switch between several accounts on one browser, "Remember this device" now applies to each account separately.
• Fixed: You can drag the AI Chat pill out of the way, and it remembers its position between visits.
• Fixed: When you reach your daily free-tier limit, background tasks now stop cleanly with a "daily limit reached" message.
• Fixed: Changes to your own email address will now be recorded in the user auth audit log.
• Fixed: Desktop users may now see an approval card when the agent runs a script it changed with a shell edit.

Manage your worktrees and subscriptions more effectively

Organizations can now manage worktrees and view subscription details more clearly. A new guide for connecting model servers is also available. Additionally, several fixes improve tool functionality and display.

• New: Settings now has a Worktrees page that lists and allows removal of unused worktrees.
• Improved: Organizations can now see mirrored price, currency, and cancel state information on subscriptions.
• New: A new guide for connecting model servers has been added to the Desktop settings.
• Fixed: Fixes a rare failure where DeepSeek tool-call history could end in an error instead of an answer.
• Fixed: Veo 3.1 Fast is no longer available for new videos; new videos now use Gemini Omni Flash.
• Fixed: The web client now correctly displays empty fenced code blocks instead of 'undefined'.
• New: Organizations can now serve local model tools without needing a backend.
• New: The Models screen can now store a Hugging Face token for accessing private and gated repositories.

Service worker installation issues resolved for returning browsers

Returning browsers can now successfully install a new service worker without issues caused by unnecessary PDF.js data files. This fix improves the overall user experience when updating the service worker.

• Fixed: Returning browsers can now successfully install a new service worker without issues caused by unnecessary files.

Remove access to files when sharing is revoked

When sharing is stopped, files are removed from the recipient's system files and conversations. Unavailable files can now be removed from system prompts. Project owners no longer retain access to files from departing members. Organizations can now load the Cloudflare Turnstile script for email sign-in and OAuth logins.

• Fixed: When sharing is stopped, files are removed from the recipient's system files and conversations.
• Fixed: Unavailable files can now be removed from system prompts to stop warning messages.
• Fixed: Project owners no longer retain access to files from departing members.
• Improved: Organizations can now load the Cloudflare Turnstile script for email sign-in and OAuth logins.

Control custom connectors from a new organization settings tab

Organizations can now manage custom connectors through a new Connectors tab, allowing control over server access and approval requirements. The multi-factor authentication setup window is now scrollable on smaller screens. The agent authoring assistants are available via the public API. Desktop conversations can now pass messages through up to five sessions. Organizations can view active grants and audit events in the Permissions settings. Alarm messages in Slack are now better styled for urgency and scanning.

• New: Organizations can now control custom connectors from a new Connectors tab.
• Fixed: The multi-factor authentication setup window now fits on smaller screens and scrolls.
• New: The agent authoring assistants are now available through the public API.
• New: Desktop conversations can now pass messages through up to five sessions.
• New: Organizations can now view active grants and audit events in the Permissions settings.
• Improved: Alarm messages in Slack are now better styled for urgency and scanning.

Chat with model servers on other machines

Organizations can now chat with Ollama and OpenAI-compatible servers on other machines. Self-hosted deployments can enforce Cloudflare Turnstile on sign-in. A request error on Bedrock Claude models has been fixed, and empty fenced code blocks now render correctly in Desktop chat replies.

• New: Organizations can now chat with Ollama and OpenAI-compatible servers on other machines through the model servers settings.
• Fixed: Self-hosted deployments can now enforce Cloudflare Turnstile on sign-in when the secret key is set.
• Fixed: Fixes a request error that could occur on Bedrock Claude models when a conversation's tool history was cut off at the start.
• Fixed: An empty fenced code block in Desktop chat replies now renders correctly as an empty code block instead of the word 'undefined'.

Clear sidebar icons when dismissing pull requests

Removing a pull request from a desktop session now also clears its icon from the sidebar. This improves the user interface by ensuring consistency.

• Fixed: Removing a pull request from a desktop session now also removes its icon from the sidebar.

Share multi-file skills and bundles in the Skill Market

Organizations can now share multi-file skills and group them into bundles in the Skill Market. The Skill Market API also returns Claude Code tools and supporting files. Additionally, on machines without a GPU, Auto limits local model loading to a maximum of 32K context.

• New: Organizations can now share multi-file skills through the Skill Market, including helper scripts and reference documents.
• New: Organizations can now group several approved skills into a bundle for easier installation in the Skill Market.
• New: The Skill Market API now returns a skill's Claude Code tool list and its supporting files with the listing details.
• Fixed: On machines without a GPU, Auto now limits local model loading to a maximum of 32K context to improve performance.

Unlock starter credits without a card via GitHub verification

Organizations can now unlock starter credits by verifying a GitHub account. The Skill Market has improved instructions for Claude Code setup. Users can now manage Skill Market listings with an API key. Additional fixes enhance error reporting and user experience.

• New: Organizations can now unlock starter credits without a card by verifying a GitHub account on the Billing screen.
• Improved: Skill authors can now access improved instructions for the Claude Code team marketplace setup.
• Improved: Organizations can now publish and manage Skill Market listings using an API key with the skills:write scope.
• Fixed: The API now clearly reports when an AI model exhausts its response budget instead of failing generically.
• Fixed: Users can now sign out and sign in again if they are stuck on the enforced MFA screen.

Detect suspicious signups with burst detection and domain holds

Suspicious signup bursts can now pause pending starter credits until their email domain is reviewed. Local models can read GGUF metadata from files and Hugging Face. Users will receive rewards when invited friends provide their first answer. Live Artifacts thumbnails now show arrows and symbols. Agents can be created and edited with any selectable model. Cached signed URLs for files now expire correctly. Chat API requests with output limits below 16 tokens are now rejected. The Turnstile challenge will now run before OAuth sign-in.

• New: Suspicious signup bursts can now pause pending starter credits until their email domain is reviewed.
• New: Local models can now read GGUF metadata from files and Hugging Face.
• New: Inviters will now receive rewards when their invited friends provide their first answer.
• Fixed: Live Artifacts thumbnails now show arrows and common symbols that were previously blank.
• Fixed: Agents can now be created and edited with any model that is selectable in chat.
• Fixed: Cached signed URLs for files now expire correctly based on their signing credentials.
• Fixed: Chat API requests that set an output limit below 16 tokens are now rejected with a validation error.
• New: The Turnstile challenge will now run before starting OAuth sign-in.

New signup credits visibility and eligibility checks for enrolled accounts

Enrolled accounts can now view their released signup credits and remaining grants in Billing. The notebook actions menu fits within the viewport. The users list can group accounts by email domain and highlight unusual signup activity. Connected Drive folders now sync automatically every hour. Admin writes now use a unified confirmation dialog. CLI users can find and switch local models directly from the command line. CloudWatch alarms now post messages to a dedicated Slack channel. The admin AI chat starts minimized for better accessibility.

• New: Enrolled accounts can see released signup credits and their remaining grant in Billing.
• Fixed: The notebook actions menu in the sidebar now fits on screen and scrolls for easy access.
• New: The users list can group accounts by email domain and highlight unusual signup activity.
• New: Connected Drive folders now sync automatically every hour without manual intervention.
• New: Admin writes now use a unified confirmation dialog that includes a reason for the action.
• New: CLI users can now find, download, and switch local models directly from the command line.
• New: CloudWatch alarms now post plain-English messages to a dedicated Slack channel.
• Fixed: The admin AI chat now starts minimized and can be easily accessed without covering the page.
• Fixed: Desktop now explains when a local runtime or model is missing and provides guidance.
• Fixed: The environment banner and sidebar scrollbar have been improved for a better user experience.

Publish multi-file React apps as shareable pages

Organizations can now publish multi-file React artifacts as a single shareable page. The Completions API documentation has been refreshed to match the live spec. The export function for knowledge files now correctly downloads the user's files instead of an empty archive. Credit handling has been improved for chat overshoots and text refinements. Invite codes can now be redeemed during email signup, and new rate rules have been added for specific API routes.

• New: Organizations can now publish multi-file React artifacts as a single shareable page.
• Improved: The developer docs for the Completions API now match the live API specifications.
• Fixed: "Export All" now correctly downloads your knowledge files instead of an empty archive.
• Fixed: If a chat request costs more than your credits, the difference is now taken from your next purchase.
• Fixed: Text refinements now use credits when the model runs, aligning with chat usage.
• Fixed: Invite codes can now be redeemed during email signup on invite-only instances.
• New: New rate-based rules have been added to limit requests for specific API routes.

Publish multi-file React apps as shareable pages

Organizations can now publish multi-file React artifacts as a single shareable page. Credit enforcement for hosted deployments is now available. Invite codes can be redeemed during email signup on invite-only instances. Other fixes include improved credit handling and API response formatting.

• New: A React app split across several files can now be published as a single shareable page.
• New: Organizations can now enforce credit limits for hosted deployments.
• Fixed: Invite codes can now be entered during email signup on invite-only instances.
• Fixed: If a chat request exceeds your credits, the difference is now taken from your next purchase.
• Fixed: Text refinements now use credits when the model runs, similar to chat.
• Fixed: Operational AI usage will now be debited from free-tier users when credits are enforced.
• Fixed: The agent-executions API no longer includes empty reasoning markers at the start of answers.

Apply rate-based rules for sign-in code and credits routes

Organizations can now apply new rate-based rules for managing traffic on sign-in code and credits routes. Additionally, free signup credits are now granted only once per inbox, even when using different email aliases.

• New: Organizations can now apply new rate-based rules for sign-in code and credits routes to manage traffic more effectively.
• Fixed: Organizations can now ensure that free signup credits are granted only once per inbox, even when using different email aliases.

Clean up worktrees when deleting conversations

When deleting a desktop conversation, unused worktrees and branches can now be removed, with clear explanations provided. Additionally, client-caused errors now return clearer 400/403 responses instead of a generic 500 error.

• New: When deleting a desktop conversation, you can also remove its unused worktree and branch.
• Fixed: Client-caused errors now return clearer 400/403 responses instead of a generic 500 error.

New email sign-in challenge option enhances security

An optional Cloudflare Turnstile challenge can now be added to the email sign-in code request process. Generated files now include information about their origins and requests. Image edits with GPT Image models now work on large source images. New accounts must verify their email before creating API keys, and may hold one personal API key until they make a purchase or the account is a week old.

• New: An optional Cloudflare Turnstile challenge can now be added to the email sign-in code request process.
• New: Generated files now include information about their origins and the requests that created them.
• Improved: Image edits with GPT Image models now work on large source images, including images you generated earlier.
• Fixed: New accounts must verify their email before creating API keys and may hold one personal API key until conditions are met.

Seven FLUX.2 image models now available for generation

Organizations can now access seven FLUX.2 models for image generation, each with distinct quality and pricing. The admin user view has been enhanced with a full-page layout for better organization. The CLI and desktop app will now reconnect to an HTTP MCP server if an idle session drops, ensuring tools remain functional.

• New: Seven FLUX.2 models are now available in image generation, each with its own quality and price.
• Improved: The admin user view has been enhanced with a full-page layout for better organization.
• Fixed: The CLI and desktop app now reconnect to an HTTP MCP server that ended an idle session.

Manage storage limits with user quota overrides

Storage limits now display their source and can be reset to default allocations. Custom limits set during account creation are preserved.

• New: Storage limits now show their source and can be reset to the default allocation.
• New: Custom limits entered when creating an account are preserved.

Limit of communication channels per account increased to 100

Each account can now have up to 100 communication channels. Analytics now includes optional fields for lane and synthetic user counts.

• Fixed: Each account can now have up to 100 communication channels.
• New: Analytics now includes optional fields for lane and synthetic user counts.

Count archived page versions toward your storage limit

Archived versions of published pages now count toward your storage limit. The app also correctly renders multi-file React artifacts in the chat preview and side panel. Moderated users will be blocked from accessing AI features.

• Fixed: Archived versions of a published page now count toward your storage limit.
• New: The app now correctly renders multi-file React artifacts in the chat preview and side panel.
• Fixed: Moderated users will now be blocked from accessing AI features.

Add offline local history and models for Desktop conversations

Organizations can now opt-in to use offline local history and models for Desktop conversations. Publishing has a new daily allowance and revisions count toward storage. The Live Artifacts section has been improved with a grid view and better visibility for access badges. Several fixes enhance the user experience, including clearer error messages and improved handling of stored gates.

• New: Organizations can now opt-in to use offline local history and models for Desktop conversations.
• New: A new grid view with thumbnails has been added to the Live Artifacts section.
• Improved: Publishing now has a daily allowance: 200 publishes per day, or 20 for accounts in their first week.
• Fixed: Completion requests now return a clear validation error for unsupported assistant prefill.
• Fixed: The chat settings now prevent empty replies from models when the output budget is reached.
• Fixed: The Live Artifacts section now shows one access badge per row for better visibility.
• Fixed: The sharing panel now displays stored gates on non-public live artifacts for owners.
• Fixed: Owners can now clear a stored gate directly from the share modal on non-public pages.
• Fixed: The publish share modal now matches access colors to the Live Artifacts badge for clearer visibility.
• Fixed: Usernames starting with 'persona-' are reserved for test accounts and cannot be chosen.
• Fixed: When another app launches b4m with a task, it no longer waits on the folder-trust question.

Media URLs are displayed again in the public feed for What's New

The public feed for What's New will now display media URLs for notes again, including videos and images.

• Fixed: The public feed for What's New will now display media URLs for notes again, including videos and images.

Daily Slack report now ranks users by credits spent

The daily Slack activity report now ranks users based on credits spent, linking directly to their admin profiles. The limit on published pages per user and organization has been removed. In 'Approve for me' mode, one 'Allow in this chat' on a command now covers later runs with different settings. When generating audio, requests now wait for capacity instead of failing, with refunds for busy service responses.

• New: The daily Slack activity report now shows who spent the most credits, linking to each person's admin profile.
• Improved: The limit on the number of published pages per user and organization has been removed.
• Fixed: In 'Approve for me' mode, one 'Allow in this chat' now covers later runs of the same command with different settings.
• Fixed: When generating audio, requests now wait for capacity instead of failing, with refunds for busy service responses.

Show images in release notes without video

The What's New section can now display a still image for releases without a video. Additionally, the product now reports billing and account lifecycle events to Overwatch. Other fixes include improved session handling on desktop and CLI, and clearer notifications for truncated responses from Bedrock Claude.

• New: What's New can now show a picture for a release that has no video.
• New: The product now reports billing and account lifecycle events to Overwatch.
• Fixed: When your desktop session stops being accepted, the app now renews it automatically.
• Fixed: Replies that get cut off at the model's output limit now show a notice explaining the response was truncated.
• Fixed: In Agent Mode, Bob Panel now asks before it starts a site read, using your credits.
• Fixed: `b4m --allowedTools mcp__<server>__*` now auto-approves that MCP server's tools.
• Fixed: Desktop users may see an approval card when the agent runs a script after editing.
• Fixed: Self-hosted deployments can now send completions to models added to the model catalog.

Organizations can manage a blocked email domain list for the signup gate

Organizations can now manage a blocked email domain list for the signup gate. The Desktop footer displays the monthly credit allowance. The workspace panel controls have been redesigned for better usability. Self-hosted deployments now check hourly for storage usage discrepancies. The expiry time for presigned URLs is now capped at 3600 seconds. Various fixes improve user experience, including tag renaming limits and session-only features.

• New: Organizations can now manage a blocked email domain list for the signup gate.
• New: The Desktop footer now displays the monthly credit allowance.
• New: The workspace panel controls have been redesigned for better usability.
• New: Self-hosted deployments now check hourly for storage usage discrepancies.
• Fixed: The expiry time for presigned URLs is now capped at 3600 seconds.
• Fixed: Renaming a tag to a name longer than 130 characters is now refused with an error.
• Fixed: CLI sandboxed commands can no longer read sensitive credential files.
• Fixed: CLI users will see the shell permission prompt more often, now naming what it covers.
• Fixed: An admin's API key lacking the admin scope can no longer perform admin actions.
• Fixed: A newly uploaded profile photo now loads correctly after uploading.
• Fixed: The sound-effects prompt limit is now 450 characters, matching provider limits.
• Fixed: Creating a voice agent with a blank description now succeeds.
• Fixed: Fixed a second scrollbar issue in the desktop app for better navigation.
• Fixed: The in-app browser can now take screenshots of off-screen pages.
• Fixed: Requesting a text completion with non-text models now returns a clear error.
• Fixed: The application now handles empty objects in ranged S3 reads without errors.
• Fixed: The notebook and Keep relay routes are now session-only features.

Offer video generation on in-app voice calls

Organizations can now request video generation during voice calls in the app, with the clip appearing in chat once ready. Additionally, the CLI sanitizes untrusted text to enhance security, image replies no longer get stuck on 'Running...', and bundle file names with a percent sign are rejected.

• New: Organizations can now ask for a video during a voice call in the app, with the clip appearing in chat when ready.
• Fixed: The CLI now strips terminal control sequences from model output and error messages for better security.
• Fixed: Image replies no longer stay stuck on 'Running...' and now show the image or an error instead.
• Fixed: Bundle file names with a percent sign are no longer accepted.

New desktop app download page and prompt

You can now download the desktop app from a new page in the web app. The Image Generation tool shows the image model and credit price before generating. Code sessions in the desktop app now feature a floating session panel for better context management. Admin settings now display units and dollar values for credit pricing, enhancing clarity. Various fixes improve error handling and user experience.

• New: You can now download the desktop app from a new page in the web app, linked from the sidebar and the profile menu.
• New: The Image Generation tool now shows which image model it will use and roughly how many credits each image costs, before you send.
• New: Code sessions in the desktop app now show a session panel in the top-right corner with project and context details.
• New: Admin settings now display units and dollar values for credit pricing, making costs clearer.
• Fixed: Invalid completion requests now return a clear, non-retryable error instead of an internal server error.
• Fixed: Completion requests with no usable user or assistant message now return a clear validation error instead of failing.
• Fixed: When a model takes a long time to respond, chat now says it is slow and retrying instead of that the service is busy.
• Fixed: Bike4Mind Desktop now shows who approved or declined each action, and approval requests wait for your answer instead of expiring.
• Fixed: A published bundle can now contain several linked pages behind one passphrase, with videos and PDFs working on protected pages.
• Fixed: Billing routes now properly check model admission and reserve credits for usage.

Amazon Nova Pro is now available as a chat model

Amazon Nova Pro is now accessible as a chat model. The desktop app's dark mode aligns with the web app's color palette. API callers using Claude models can now defer tool definitions. The Bike4Mind CLI allows for message queuing and cancellation while typing. The skills list in Customize is now more concise, with collapsible skill groups. The desktop agent now lists all installed skills. A bearer token is now required for specific CLI commands.

• New: Amazon Nova Pro is now available as a chat model.
• Improved: The desktop app's dark mode now uses the same blue-tinted colors as the web app.
• New: API callers using Claude models can now defer tool definitions and let the model search for tools.
• New: In the Bike4Mind CLI, you can queue messages while the agent works.
• Improved: The skills list in Customize is shorter and easier to scan with collapsible skill groups.
• Fixed: The Create PR button is now more compact, and queued pull requests show an amber icon.
• Fixed: The desktop app now finds Nix-installed command-line tools when launched from Finder or the Dock.
• New: The desktop agent now sees every skill you have installed.
• New: A bearer token is now required for the command `b4m mcp serve --http`.

Record timing and retries for chat replies

Users can now see timing and retry information for chat replies. New accounts experience a quieter signup session. The desktop app shows a notice for new versions.

• New: When a reply is slow or had to retry, a note under it indicates so, with a turn ID for support.
• New: New accounts get a quieter first session without announcements interrupting setup.
• New: The desktop app now shows a small notice when a newer version is available, with a link to its release notes.

Add a copy and save menu to chat images

Users can now right-click on images in conversations to copy, save, or attach them. Additionally, generated videos can be delivered to the requesting Slack thread. The desktop app introduces a new preference for starting Code sessions in their own git worktree. Other improvements include a cleaner login page and cookie banner, and the ability to rename sidebar sections. The Artifacts page now displays items as a thumbnail grid. Organizations can also attach videos to release notes and see a new consent screen for federated admin checks.

• New: Right-click an image in a conversation to copy it, save it, or attach it to your next message.
• New: You can now ask the Bike4Mind Slack bot to make a video; it posts the finished clip back in the same thread.
• New: The desktop app has a new 'Worktree by default' setting that starts new Code sessions in their own git worktree automatically.
• Fixed: Fixed Desktop failing to find GitHub CLI and other command-line tools installed through the user's shell environment.
• Fixed: The sign-in page and cookie banner are cleaner with less repeated text and improved error handling.
• New: In the desktop app, you can now rename a sidebar section by double-clicking its name or right-clicking it.
• New: The desktop Artifacts page now shows your artifacts as a grid of thumbnails, making them easier to browse.
• New: Organizations can now attach short videos to each What's New release note after a production release.
• New: Organizations can now see a new consent screen that lists the federated:admin-status scope during the federated admin check.
• Fixed: Organizations can now only open Image and Productions features when Create v2 is enabled.

Manage OAuth clients from an admin page

Bike4Mind admins can now create, edit, deactivate and rotate Sign in with Bike4Mind OAuth clients from an admin page instead of a script.

Bob sidebar row uses its own icon

The /bob row in the Bike4Mind sidebar now draws the icon its nav item carries instead of a stock glyph.

Data lake Create wizard starts from the source

The Bike4Mind data lake Create wizard now opens on Where's your content, with cards for Upload files, Google Drive and GitHub repository.

More reference images for GPT image models

Bike4Mind raises the reference-image cap for GPT image models to 15.

Bob no longer overstates what it checked

Bob now checks shipping, returns and reviews before reporting them absent, and tells readers a page's location line shows where Bob fetched it, not where they are.

Bob landing share card and llms-full.txt

The Bob landing page gains a share card for link previews, a touch icon and an llms-full.txt file.

Claim an unseeded Fleet from Setup

A new company can claim an unseeded Fleet from the Setup page after signing in, without a developer running the seed script.

K2 404 page links to every app

Unmatched K2 URLs now show a Page not found page with links to every K2 app.

Paintbox completes its 27 blend modes

Paintbox adds the remaining blend modes, including Linear Burn, Vivid Light, Hue, Luminosity and a seeded Dissolve, for 27 in total.

Paintbox develop presets

Paintbox adds six built-in develop presets, including Warm Film and Cool Matte, and an apply_preset agent tool that lands a preset as a group.

Paintbox Ink Cascade icon

Paintbox replaces its favicon with the Ink Cascade icon, three stacked ink sheets in yellow, magenta and cyan.

Paintbox MCP server renamed

The Paintbox MCP server id is now paintbox instead of paint-box, so existing clients need to re-add it.

Bob landing page new look

The Bob landing page at bob.bike4mind.com has a new look and the Bob logo, and keeps its light mode.

K2 shares daily accuracy as a percentage bar

K2 result shares now show the exact daily accuracy percentage and a 10-block green and red bar, replacing the emoji grid.

Share dialog scrolls and shares text

The share dialog now scrolls on short screens instead of clipping the card, and shows the share text with a Copy text action.

K2 Tempo shows the game first

K2 Tempo Notes and Chords now show the drill above the achievements, and the header controls wrap inside mobile screens.

Companion View terminal is closed by default

The Manifold Companion View terminal is now a toggle that starts closed, so opening a companion no longer starts a CLI process in the background.

No pane limit in the Manifold footer

The Manifold footer terminal panel no longer caps how many instances a project can run, and its instance nav scrolls.

Manifold console can retry a stuck stage

A Manifold project console can now relaunch a stage stuck in error, action required or idle through a new project_retry_stage tool.

Overwatch counts unique users per event

The Overwatch Events per day card has a Count picker for events, unique users per day, or unique users over a rolling 7 or 30 days.

K2 tags CJK text with its language

K2 now tags Japanese, Chinese and Korean text with its language, so browsers pick the right Han glyph shapes and screen readers pick the right voice.

Overwatch product rows show their own events

Expanding a product row in the Overwatch Products table now shows its first-party events next to the product detail.

New companions show in the drawer at once

A newly created Manifold companion now appears in the Companions drawer immediately instead of staying hidden after the drawer is reopened.

Choose where Manifold keeps its data

Manifold Settings has a Storage section for moving its data folder to any folder you pick, such as an external drive, after a checklist runs against the chosen folder.

Backtests corrected for stock splits

Stocks and Vibes backtests now use 20 years of price history corrected for forward and reverse stock splits.

Manifold checkpoint commits pass CLA checks

Manifold checkpoint commits are now authored by the repo's configured git identity, so they no longer leave CLA checks red on the pull requests they land in.

Manifold scope hooks

Manifold adds scope hooks, authored as hooks files at the column, board, project or global tier and loaded on both the claude and opencode engines.

Manifold teardown and send-back fixes

Proceeding into a teardown column no longer hangs while the worktree is deleted, and declining a send-back with Keep here now parks the ticket as idle.

Overwatch chart of any first-party event per day

Overwatch adds an Events per day card that charts any first-party event per UTC day, optionally split by one metadata field.

Run Manifold claude stages without auto-memory

A Manifold column can start its claude stages with Claude Code auto-memory turned off, using a new checkbox in column settings.

Manifold ticket with a dead stage lands in error

A Manifold ticket whose stage engine exits unexpectedly now lands in error instead of showing processing forever.

JavaScript in Manifold HTML artifact previews

Manifold HTML artifact previews can run JavaScript through an off-by-default setting in Appearance, behind a one-time risk notice.

Manifold Launchpad live status

Manifold Project Settings now shows live Launchpad status: a port chip for each local server, an AWS profile session dot and a present or missing mark for each tracked secret.

Tarot starts the reading at the deal

For a signed-in sitter with a shuffled deck, Tarot now offers Deal and read, which starts the reading as the cards fall, while Deal only still deals for free.

Tarot deep reads on Opus 4.6

Tarot deep reads now run on Opus 4.6, and the deck chooses the reader persona: Deco gets Madame Odile and Classic gets Miss Honora.

Tarot readings no longer come back empty

Tarot fixes readings that showed no narration at all, and now gives an aside per card before the whole narrative.

Tarot crescent-and-star favicon

Tarot gets a gold crescent-and-star favicon on the dark green of the table, with a home-screen icon.