Open a page in the browser this MCP session owns and return its indexed action space. The session reuses that browser and its logins across calls; the previous tab is closed first.

Read the current page again: the same indexed action space under a new observation number. Use it after the page changes on its own, or after a refused action.

Run one operation on one target from an observation: CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT. CLICK, TYPE_TEXT, and SELECT take a target index from that observation; TYPE_TEXT also takes the text. SCROLL_UP, SCROLL_DOWN, and WAIT take no target. You may instead omit operation and target and give an intent: the policy chooses the operation and target on that observation and runs it. A TYPE_TEXT with an intent and no text gets its value from the text model. Both spend model quota; --jev-decider says whether Jev or the text model answers.

Evaluate one JavaScript expression in the page this session owns and return what it produced: a string as itself, any other value as JSON, capped so one dump cannot crowd out the conversation. Use it for a programmatic read the indexed action space cannot express, or a one-off DOM operation. It runs on this session's one CDP connection: never start a second client — a raw websocket, a puppeteer/playwright script, another browser-harness daemon — because Chrome 144+ asks "Allow remote debugging?" once per connection and two drivers race each other on the same page. The expression gets no user gesture, a thrown error comes back as text, and it produces no observation: call browser_page before the next browser_act.

Capture the visible viewport as an image. Target boxes and page structure are not drawn into it.

Read what the page reported since the last read: console messages, uncaught exceptions, and failed or 4xx/5xx requests. Each read drains the buffer, so read it after reproducing a problem.

Hand one goal to the browser policy in its own tab and return the trace it produced. The policy chooses operations and targets by itself and spends model quota — --jev-decider says whether Jev or the text model answers; the page this session already has open is left untouched.

Close the tab this session owns and stop its daemon. The browser this session attached to is never closed: other sessions may still be driving it.

Report what the browser tools need and whether it is there: the Python interpreter, the engine and browser-harness versions, the managed environment, the browser that would be launched, the attach target, the configuration in use, and the fix for anything missing. Call it first when browser work fails. With install: true it builds a missing engine environment first and waits for that install.