The browser is an MCP server
skep does not ship a browser stack. Browser automation arrives the same way every external capability does: an MCP server the operator registers, whose tools decide under the one policy engine (invariant I5). The reference server is Playwright's official one.
Register it
From chat (the Queen proposes, you confirm the card), or ask for exactly:
register_mcp_server
server_id: playwright
transport: stdio
command: ["npx", "@playwright/mcp@latest"]
scope: browse
Requirements: npx on PATH (npm ships it). skep doctor warns if a
registered browse server's launcher is missing. No npm? Run the server any
other way that yields a stdio argv (a container wrapper works: podman on
this machine) and register that argv instead.
What the browse scope means
Tools on a browse-bound server decide as browse/read or browse/act
(v71-F2, mirroring the v41-F3 email precedent):
- read — page-STATE reads:
snapshot,screenshot,console,network_request,wait_forshaped names that the server itself marks read-only (readOnlyHintin its tools/list) and that carry no act verb (submit,click,fill, …) anywhere in the name (ES-F54). skep asks the server's own tools/list for the hint once per registration (ES-F160: registering the server again, or restarting serve, asks afresh). A Playwright server (@playwright/mcp) that sends no hint for a tool falls back to skep's list of its page-state reads (ES-F161):browser_snapshot,browser_take_screenshot,browser_console_messages,browser_network_requests,browser_network_request,browser_wait_for; a hint the server does send always wins. Any other server that does not send it, or a list that fails, makes every one of its tools an act (a card). These run inside the turn unless a policy rule asks for them: the builder and operator baselines carrybrowse-read-gated(ask), the assistant baselinebrowse-read(allow), and a chat's role tightens the same way (ADR 0059 P3-3). - act — everything else, navigation included:
navigate,click,type,fill,press,select,evaluate,tabs, uploads. Each cards until you allow it. Navigation cards deliberately — it is an outbound fetch, andread_urlset the parity: every fetch is one operator decision.
Each call starts the server, waits for its answer, then stops it (one
process per call). A page does not carry over from one call to the next:
browser_navigate answers with the snapshot of the page it opened, and a
later browser_snapshot starts on about:blank. The server runs in
~/.skep/browser/, never in the folder serve runs from: Playwright writes
its snapshot and console files to .playwright-mcp/ there, and a read's
optional filename can only land inside that folder.
Teach it your posture with allow_mcp_tool (writes a learned
browse-scope rule; a policy deny always wins and refuses without a
card), e.g. allow playwright:browser_navigate once you trust the flow.
The walls that do not move
- The Queen still never holds the trigger: an unallowed act is a card, never a default (I6).
- Workers get nothing from this: the browse scope is a Queen-side surface; worker capabilities and the git guards (I4) are untouched.
- Forged tools (the v71-F1 forge) may NOT shell out to a browser themselves — a tool that needs the web goes through this same door.