Skip to content

Browser DOM and DevTools automation through MCP

Ru has no built-in browser tool. web_fetch returns page text, and computer works on screenshots, UI Automation elements and synthetic input. To let the Agent read a page's DOM, console and network log, or click and type by element, attach a browser automation MCP server. It is reached through the existing mcp tool (ADR 0025, MCP servers), so this adds no tool definition, no manifest change and no new execution path.

This page is a recipe, not a shipped default: nothing here is enabled until you edit your own %USERPROFILE%\.ru\mcp.json. The servers are third-party programs that need Node.js; Ru does not bundle or update them. The configurations below follow the servers' published options and were not run end to end in this repository.

Playwright MCP with the installed Microsoft Edge

@playwright/mcp needs Node.js 18 or newer. With --browser msedge it drives the Edge that ships with Windows instead of downloading a browser.

{"servers":{"browser":{"command":"npx.cmd",
  "args":["-y","@playwright/mcp@latest","--browser","msedge","--isolated","--headless",
          "--image-responses","omit"],
  "description":"Isolated Edge: page snapshot, console, network, click and type by element",
  "timeout_seconds":120}}}
  • --isolated keeps the profile in memory. Without it the server uses a persistent profile on disk; either way it is not your own Edge profile, so the Agent has none of your cookies or sign-ins.
  • --headless hides the window. Remove it to watch the browser.
  • --image-responses omit matches what Ru can return: MCP results carry text, structured content and resource links, and images are counted but not returned. The Agent verifies a page with browser_snapshot, browser_console_messages and browser_network_requests, not with screenshots.
  • Replace @latest with a version you have reviewed. npx -y otherwise runs whatever the registry serves at the next start.
  • The Broker bounds server startup to 20 seconds. The first npx -y run downloads the package and can exceed that (mcp_server_unavailable); run the same command once in a terminal, or install the package globally and point command at it.

Chrome DevTools MCP

chrome-devtools-mcp drives Google Chrome (current stable or newer) and adds performance traces and Lighthouse audits. It needs a Node.js LTS release and an installed Chrome.

{"servers":{"devtools":{"command":"npx.cmd",
  "args":["-y","chrome-devtools-mcp@latest","--isolated","--headless",
          "--no-usage-statistics","--no-performance-crux"],
  "description":"Isolated Chrome: DOM snapshot, console, network, performance",
  "timeout_seconds":120}}}

--no-usage-statistics and --no-performance-crux stop the server from sending usage data and trace URLs to Google. Do not use --autoConnect, --browserUrl or --wsEndpoint here: they attach to a browser that is already running, which may be your own profile.

What the Agent can do, and what approves it

plan denies every MCP call. In standard each call asks, and the card names the server and the tool; full_access does not ask. Subagents never receive the mcp tool.

"Always allow" saves a rule for one server and one tool, which is how to get a usable flow without giving up the prompts that matter:

Tools Suggested handling in standard
Playwright browser_snapshot, browser_console_messages, browser_network_requests; DevTools take_snapshot, list_console_messages, list_network_requests Always allow: they only read the current page
Playwright browser_navigate, browser_click, browser_type, browser_fill_form; DevTools navigate_page, new_page, click, fill Ask each time, or always allow once you trust the task
Playwright browser_evaluate, browser_run_code_unsafe; DevTools evaluate_script Keep asking: these run arbitrary JavaScript in the page

Boundaries to know before enabling it

  • Navigation is not scoped per site. The approval card shows the server and tool, not the arguments, so approving browser_navigate does not show the URL, and an "Always allow" rule covers every URL. The arguments are still bound to the single-use approval and recorded in the run.
  • WebPolicy does not apply. web_fetch refuses loopback and private addresses (web_private_host). A browser started by an MCP server is a separate program running as you: it can open http://localhost:… (which is what makes it useful for a dev server) and also any other address your machine can reach. Ru does not confine what a server does once it runs.
  • Origin lists are advisory. Playwright's --allowed-origins and --blocked-origins are documented by its authors as not a security boundary and not applied to redirects. They reduce accidents; they do not contain a hostile page. Chrome DevTools MCP's --allowedUrlPattern/--blockedUrlPattern are likewise the server's own check.
  • Page content is untrusted. MCP results are marked untrusted like web content. A page can contain text addressed to the Agent; it grants no authority, and the approval prompts above are what stand between such text and an action.
  • Configuration stays yours. mcp.json is read only from your user profile and is a protected path, so the Agent cannot add or change a server in plan or standard.

A built-in browser tool with per-origin grants and WebPolicy on every request would close the first three gaps. It needs its own ADR and a tool-manifest change and is not planned here.