Skip to content Browser Use for Every Model, Timeline Chat, and Live Steering
Features
- In-app browser for every model — Agents drive the browser in the side panel the way Codex does, with the same tools on Claude, GPT/Codex, Muse and every other model. A persistent, sandboxed JavaScript runtime gives the agent the page's accessibility tree, clicks and typing, Playwright-style locators, coordinate input, screenshots, dialogs, the clipboard and dev logs. Browsing runs in the background, so agent tabs keep working when the panel is closed or you switch chats. Open the panel to watch the agent's cursor at work, or ask it to show you its tab. Screenshots reach every vision-capable model and show as thumbnails in the chat, and each browser step shows the site's icon. The previous
browser_tool is still available in Settings → App → Tools. - Browser safety and control — The first time an agent acts on a new site in a session, you choose Allow for this session, Always allow, or Deny. Local dev servers never ask, and Settings → App → Browser Site Access keeps your always-allowed and blocked sites. File uploads, and downloads the agent starts, ask first unless you allow them. Click or type in a tab the agent is driving and you take over for the rest of the turn.
- Annotate pages — Turn on Annotate in the browser toolbar, click an element or drag across an area to drop numbered comment pins, then send them to the composer as text plus cropped screenshots. Hold to view the page without the pins.
- Browser mentions — Type @ in the composer to point the agent at @Browser or at one of the chat's open tabs.
- Live steering — Messages you send while an agent is working now steer the running turn on Claude, Cursor and Pi models instead of interrupting it. They reach the agent at its next tool boundary, keep their order, and anything not yet delivered goes back to the queue instead of disappearing or falsely reporting success.
- Timeline turn cards — Each turn is one timeline: a live "Working" header, then tool groups, plan steps and questions as nodes on a single line. Code Mode scripts from GPT and Muse models read as "Ran script" rows with highlighted source and readable output.
- Subagent tray — A tray above the composer lists every agent in the session with its status, model and current action. Dismiss it once no agent is working; it returns when one starts again.
- Redesigned question card — Questions from the agent show numbered options with a Recommended tag and keyboard shortcuts: 1–N to pick, Enter to submit, Esc to skip.
- New model picker and usage warnings — The model picker groups providers in one popover, context usage shows as a ring with a compaction popover, and a banner warns at 80% and 95% of a 5-hour or weekly limit.
- Streamable HTTP for MCP sources — MCP sources connect over Streamable HTTP and fall back to SSE only while connecting.
- Muse Code usage — The usage card shows the 5-hour and weekly windows of a Muse Code membership.
Improvements
- Faster chat while streaming — Finished turns no longer re-render on every streamed update, idle animations stop repainting every frame, and the Review tab no longer slows the app while a chat streams.
- Smaller debug logs — The network debug log rotates by size instead of growing to hundreds of megabytes.
Bug Fixes
- No leaked MCP connections — Switching models or providers, reverting, retrying sign-in or deleting a ChatGPT/Codex, OpenAI, custom-endpoint or Cursor session now closes its MCP connections and running tool commands instead of leaving them open until quit.
- Subagent model labels reflect the run — Agent detail headers and roster badges show the model the child actually used, not an outdated profile selection.
- Agent-name hover shows the full brief — Spawned agents show their original assigned directions on hover, including older runs whose parent task only stored the short agent name.
- Muse effort and usage — Standard Muse models keep XHigh distinct from Max effort, and the usage card keeps its last good meters when Meta briefly omits quota data.
- Chat polish — The chat header shows only the title, the Edited-files card waits until the agent is completely done, brand icons are back on the timeline, and the timeline line no longer breaks between events.