What’s new

Todo board, in Experiments

Turn on Todo board in the new Settings → Experiments tab, and Todo appears in your sidebar. Draft tasks, hand each one to an agent, and see what's working, what needs you, and what's done. Experiments are off by default and set per person.

MiniMax M3.1 Flash Preview

MiniMax M3.1 Flash Preview is now the default model on MiniMax Subscription, with a 1M-token context window and image input on both Token Plan and M Plan. Agents set to the latest MiniMax switch over, and M3 stays selectable. Effort is fixed at high.

H3 video on MiniMax

MiniMax M Plan sessions can now generate video with MiniMax H3 and H3-Max, drawing on your plan's video quota. Agents try H3 first and fall back to Hailuo 2.3 on older Token Plan seats that don't include it.

GLM-5.3 Standard

GLM-5.3 Standard joins GLM-5.3 Flash in the Z.AI tab on Platform. It's the full GLM-5.3 model, text-only, with a 1M-token context window and low, medium, and high effort, at $1.40/$4.40 per million input/output tokens. Flash stays the default.

Fixes and improvements

• Grok subscription sessions work again and adapt when xAI raises its minimum version.
• GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna can now run as subagents.
• Agents set to the latest GPT now use the cheaper GPT-6.1 Sol.
• Long Codex subscription sessions reuse the prompt cache on nearly every turn.
• Agent-to-agent file transfers now work on cloud workspaces.
• Provider errors show the vendor's own message instead of "Upstream error".
• The model picker lists GPT by tier: Astra, Sol, Terra, and Luna.
• GPT-5.4 and GPT-5.5 are no longer offered on Platform or Codex.
• Session menu items show a spinner while a fork or raw log runs.

GPT-6.1 Sol

GPT-6.1 Sol is now available in the Codex subscription picker and on Platform, where it costs $2/$10 per million input/output tokens, rising to $4/$15 past 272K tokens of context. Agents set to the latest GPT stay on GPT-5.6 Sol.

Diagrams in chat

When an agent answers with a Mermaid diagram, the chat now draws it instead of showing the code. Diagrams follow light and dark mode. Flowchart, sequence, state, class, ER, and XY charts render, while Gantt, pie, and other types stay as code.

HighLevel connected accounts

Connect a HighLevel business as a connected account, and your agents can work in it through HighLevel's API. Each connection covers one business and is named after it. Agency logins aren't supported, so sign in as a user of one business.

Keyboard playback in the file drawer

Video and audio in the file drawer now take YouTube-style shortcuts and play from 0.25× to 2×, and video steps one frame at a time. Press C to comment at the current frame, or M to speak one. Shortcuts stay off while you type.

Fixes and improvements

• Voice mode started from the agent home now works in Safari.
• Always allow now saves for remote MCP server requests.
• The Stop dialog now matches the background tasks actually running.
• Workflow previews fill in live, with agents under the right phase.
• Chat rows and request cards wrap neatly beside an open file.
• Agents see the real address when a browser page fails to load.
• Drawer comment times now include hundredths of a second.
• In drawer comments, Enter adds and Shift+Enter starts a new line.
• The drawer's comment box now stays inside the drawer.
• Voice transcripts read left to right, with clearer speaker turns.

Claude Sonnet 5.5

You can now use Claude Sonnet 5.5 on Anthropic, Claude subscription, OpenRouter, and Platform connections, with all five effort levels. It becomes the default Sonnet at the same $2/$10 per million input/output tokens as Sonnet 5. Sessions pinned to Sonnet 5 stay put, and Bedrock doesn't offer it yet.

MiniMax Subscription connections

Add MiniMax Subscription in Settings → Model Providers, sign in with a device code, and run agent sessions on your MiniMax Token Plan with MiniMax M3. Choose the Global or China region at sign-in. Usage bars track your plan's 5-hour and weekly limits.

Images, speech, and video on MiniMax

You can now generate images, speech, and short videos in MiniMax Subscription sessions with Hailuo 2.3, drawing on your Token Plan allowance. Results are saved to the agent's workspace. Video depends on your plan tier, and it only works while the session runs on MiniMax.

Fixes and improvements

• Local Chrome now keeps logins made right before the browser closes.
• Old thinking cards no longer reappear after a follow-up during background work.
• Stopped sessions now release their memory instead of lingering.
• Replies after a background agent finishes no longer fail mid-turn.
• Agent templates now export and publish AGENTS.md instructions.

Kimi Subscription connections

Add Kimi Subscription in Settings → Model Providers, sign in with a device code, and run agent sessions on your Kimi Code membership with K3. Choose the US or China region at sign-in. Web search isn't offered, and costs shown are API-equivalent estimates.

Images and video on your subscription

Codex Subscription sessions can now generate and edit images, and Grok Subscription sessions can also make video with Grok Imagine, using your plan's allowance. Results are saved to the agent's workspace. It only works while the session runs on that subscription.

Fixes and improvements

• Shared agents now know which teammate sent each message.
• Wide drawers slide over the chat instead of crushing it.
• Codex Subscription chats no longer stall in local mode on Apple containers.
• Brainstorm with Voice on agent home now starts a voice mode session.
• The in-chat Voice feedback button has been removed.
• Agent instructions move to AGENTS.md, and existing CLAUDE.md files are renamed.
• The model picker shrinks to its icon when a drawer narrows the chat.

Codex Subscription connections

Add Codex Subscription in Settings → Model Providers with a device code sign-in, then run agent sessions on your ChatGPT plan. Codex models offer a Fast speed, which uses your plan's credits faster. Summaries still need an API-capable connection.

Grok Subscription connections

Sign in to a Grok subscription from Settings → Model Providers with a device code, then choose it for agent sessions. Image input and web search both work. Costs shown for subscription connections are API-equivalent estimates, not your plan's charges.

OpenAI-compatible endpoints

Generic connections can now use the Chat Completions or Responses API, so OpenAI-compatible endpoints work for agent sessions and summaries. Chat Completions can send max_tokens for servers that require it. Existing Generic connections keep the Messages format.

Model Providers, redesigned

Settings → Model Providers lists your connections together and adds new ones through a guided dialog: pick a provider, then follow its steps. Usage meters and reset times for Codex, Grok, and Platform show here and in the model picker. Exhausted accounts stay selectable.

A new voice mode composer

Voice mode now shares the text composer's frame, with an animated orb and a live transcript of both sides. Press Space to interrupt, or mute the agent's voice or your microphone from the controls with ⌘⇧M and ⌘⇧A.

Fixes and improvements

• Email and chat now deliver replies from models that don't stream text.
• Triggers start delivering events after you connect Platform, no restart needed.
• Agents with only an AGENTS.md use it as their instructions.
• Email inbox setup no longer fails on an expired Platform sign-in.
• Agents give their real inbox address when asked.
• Linear integrations clear their error badge after reconnecting.
• The model picker lists connections with provider logos, Platform first.
• Default connections are renamed Anthropic API and Gamut Platform.

Email inboxes for agents

Give each agent a dedicated email inbox from its integration setup. Choose who can write to it, then let the agent read messages, reply in the same thread, and send mail with attachments. Available on Platform-connected deployments; incoming mail follows the access level you choose.

Model connections for each session

Settings → Model Providers now lets administrators share global connections and members add personal ones, including multiple accounts with the same provider. Choose a connection and model per session, agent default, or automation. Shared sessions can use global connections or your own.

Claude Subscription connections

Add Claude Subscription in Settings → Model Providers with a token from claude setup-token, then choose it for agent sessions alongside your API-key connections. Browser sign-in and automatic token refresh are not available, and summaries still need an API-capable connection.

Fixes and improvements

• Accepted integration messages retry after restarts or transient failures.
• Chat shows provider cards for Slack, Telegram, iMessage, Linear, and email.
• The Local/Cloud switch is visible to every desktop user.
• Cloud workspace mobile pairing opens in your browser.
• Local Chrome recovers from Google passkey sign-in prompts.
• Duplicate analytics targets no longer blank the workspace.
• Onboarding shows available runtimes when no built-in option exists.
• Global summaries only allow API-capable model connections.
• Browser and dashboard model fields use the grouped model picker.
• The Usage link follows your selected model provider.

Connect Linear

Connect Linear from External Integrations so an agent joins as its own named app. Mentions and thread replies open in that agent, which replies in Linear as itself. Each agent needs its own OAuth app, and missed events are not replayed.

Opus 5.5

Anthropic's Opus 5.5 is in the model picker and is now the default for opus. Input and output are $4 and $20 per million tokens, with fast mode on Platform. Opus 5 stays selectable. Not on Bedrock yet.

GPT-6 Sol and Luna

OpenAI's GPT-6 Sol and Luna join Astra in the model picker on the platform provider. Sol is $2 input and $10 output per million tokens, Luna $0.10 and $0.50. Existing gpt sessions stay on GPT-5.6 Sol.

Triple Whale and TrueProfit in the MCP directory

Official Triple Whale and TrueProfit OAuth servers are now in the searchable MCP directory. Connect either like other catalog servers. TrueProfit's setup guide includes the sign-in and return steps.

Fixes and improvements

• Built-in X, audio, Exa, Apollo, and media work with any LLM.
• Recurring tasks wait for the previous run to finish.
• In-chat Connect Account and Connect MCP cards gained Cancel.
• Cancelling Account reconnect keeps you signed in to Platform.
• Connecting an MCP no longer asks about a protocol probe.
• Task lists work on Windows with WSL2.

Share files between agents

Agents can now pass files to each other, not just text. When one agent invokes another, it can send workspace attachments, and the called agent can download files the first one delivered. Every share and download needs its own one-time approval.

Grok 4.7 is available

Grok 4.7 arrives on Platform and OpenRouter, and the bare Grok option resolves to it. It keeps a 500k context window and four effort levels, at xAI's $2 input and $6 output per million tokens. Agents pinned to 4.6 or 4.5 stay there.

Fixes and improvements

• OpenAI voice mode now starts with your recent chat history in context.

Fixes and improvements

• Every model picker lists the strongest model first.
• Effort options now match what each model actually accepts.
• Voice mode no longer exits right after a billing error.
• Chat re-auth cards gained a Cancel button.
• A second sign-in attempt cancels the first window.

Fixes and improvements

• Apple Container updates delete old agent images from disk.
• Account offers Stop all for agents still on the old platform token.
• Download Raw Log in the session menu saves a .jsonl file.
• URLs in chat stay links next to fullwidth punctuation.

Live voice conversations

Voice mode on Platform now holds a live GPT-Live conversation with the selected agent, with OpenAI read-aloud and dictation. The same voice works with your own OpenAI key. Deepgram remains available if you bring your own key.

Browse and edit memories

Agent Home now has Memories, where admins can search, preview, and edit what the agent remembers. Saves check for newer edits, and leaving with unsaved changes asks you to confirm. Only agent admins can open it.

X connected accounts

Connect your X account so an agent can read your timeline, bookmarks, and DMs, and post as you. Available on Gamut's Composio, not with a personal Composio key.

Plaid connected accounts

Connect a bank through Plaid so an agent can read accounts and transactions. The connected account is named after the bank. Plaid appears only on Gamut's Composio, not with a personal Composio key.

Apollo lead enrichment

On a platform workspace, agents can look up a person or company through Apollo without your own Apollo key. They will not reveal phone numbers or write to Apollo CRM.

Fixes and improvements

• Slack mention-only threads keep follow-ups after a restart.
• Cancel appears if Account reconnect hangs in the browser.
• Voice keeps speaking when a request card opens.
• Stopping a request card also exits voice mode.
• MCP Reconnect shows only to the connection owner.

Fork & Summarize

Fork & Summarize copies a long chat into a new session and summarizes there. Use it from the session menu, or Start with a summary on the stale-conversation card. The original conversation is left alone.

Fixes and improvements

• Slack, Telegram, and iMessage recover after an agent-started message.
• Skill-spawned subagents appear in the activity panel.
• Forked chats mark the branch point in the thread.
• A sleeping cloud workspace says it's asleep, not an error code.
• Background subagents keep running after you press Stop.

Fixes and improvements

• Home graph agent lines follow agent-to-agent permissions only.
• Failed agent imports roll back instead of leaving a half-created agent.
• Template imports start without old chat history.
• Saving one secret no longer truncates another.
• Skill exports keep images and other binary files intact.

DeepSeek V4.1 Flash

DeepSeek V4.1 Flash is in the model picker on the platform provider. Input and output are $0.22 and $0.66 per million tokens, with a 1.04M context window and image input. Built-in web search and fetch are off.

Fixes and improvements

• Enter a promo code on the subscribe paywall.
• Windows installers are smaller, and auto-updates work again.
• Windows computer use shows real app names and launches Start-menu apps.
• Google Sheets, Docs, and Drive uploads request write access.
• Browser waits report how long they took, and timeouts say they timed out.
• Form fills show the field's actual value.
• Large browser results truncate instead of failing.

Fixes and improvements

• Cloud billing works when the browser blocks third-party cookies.
• Agents see each page's URL, title, and HTTP status.
• Agents can read full page text from a snapshot.
• Clicks, presses, and hovers report what changed on the page.

Profile photos

Account photos or a stable gradient show in the sidebar, sharing controls, and the typing indicator. Settings lets you upload a workspace photo or restore the account one, and the file stays after you sign in again.

Fixes and improvements

• Shared agents show member photos in the header and sidebar.
• Members can replace expired connections on a shared agent.
• Agents are told the date, weekday, and timezone.
• Agents are told when fifteen minutes have passed between messages.
• Computer use works on Windows.
• Agents load dashboard, widget, and workflow skills.

Agent widgets

Glanceable cards live on Agent Home and App Home. Agents fill them from a script, not a chat; tap a dashboard widget to open it. They refresh when you open that agent's home, after a session, or on tap — App Home never wakes containers.

Browser navigation

Back, Forward, Reload, and a read-only address sit on the in-session browser. Full screen folds the sidebar away; Escape leaves if the page didn't use it. Stop closes the browser. Navigation needs a current container image.

Fixes and improvements

• Google sign-in recovers from passkey and security-key prompts.
• Agents start after a damaged container image.
• Stop still ends a session stuck on Working.
• Viewed sessions keep title alerts cleared.
• Dictation shows an error when transcription fails.
• Failed chat follow-ups keep the last reply.
• Sidebar folder counts show only when collapsed.
• Browser tabs match the file drawer.

Voice mode

A voice button beside the mic starts a call with the agent. Speak, pause, and it sends; replies play as they arrive, and talking over them interrupts. Exit with ×. Shown only when the voice provider can both listen and speak.

Read replies aloud

Open a reply's menu and choose Read aloud. Words light up as they're spoken, with pause, stop, and speed under the bubble. Set voice and rate in Settings → Voice. Hidden when speech isn't configured.

GPT-6 Astra

OpenAI's GPT-6 Astra is in the model picker on the platform provider. Input and output are $10 and $50 per million tokens, with a 1.05M context window. It is not the default; GPT-5.6 Sol remains the latest GPT.

Stop keeps background tasks

Stop ends the reply and leaves background tasks running. When tasks are in the activity card, it asks whether to keep them or stop everything. Each running task has its own stop.

Fixes and improvements

• Dictation and voice mode transcribe in Safari.
• A typed draft survives when the agent asks for input.
• Agents are told which folders you mounted.
• Select text in file previews without the line numbers.
• Profile and login now live on the Account tab.
• Home cards keep the sessions list off the status chip.
• Cron and webhook health uses a fixed up/down pager.

Update scheduled tasks and triggers

Agents can change a scheduled task's time or prompt without replacing it, and edit trigger prompts the same way. History and linked sessions stay with the same id. There is no new screen — you ask the agent.

Fixes and improvements

• Approving a remote MCP mid-chat continues the same turn.
• Out of credit shows a paywall above the composer.
• Delivered files show as rows with name, size, and download.
• File drawer uses shrinking tabs and shows the file size.
• Uploads use the same chips in the composer and after send.
• Sleeping agents show as running when a session starts.
• Stopped turns no longer show No response requested.
• Deleting an agent also removes its connected-account triggers.
• Agents skip the current chat when searching past sessions.
• Workflow agents keep their labels and show the model.

Muse Spark 1.3

Meta's Muse Spark 1.3 is in the model picker on the platform provider. Muse Spark 1.1 and 1.2 stay selectable. Input and output are $1.25 and $4.25 per million tokens; the contributor tier may use your prompts.

Fixes and improvements

• Hidden session tabs show working, waiting, or unread.
• First chat after a cold cloud start no longer fails.
• Child agents created by another agent keep a composer.
• A spinner shows while older messages load.
• Agent home sessions show last chat time.
• Sidebar and agent home share the same menus.
• Session and API log auto-delete live on agent home.
• Export Agent from the sidebar opens the share sheet.
• Read-only skillsets cannot be chosen for Publish.
• Local Share keeps Invite as a Cloud explainer.