browser-tools-mcp

On Windows, if your client cannot find npx, use "command": "cmd" with "args": ["/c", "npx", "-y", "@agentdeskai/browser-tools-mcp@latest"].

Requires Node 22.19 or newer. Check with node --version; if you use nvm or asdf, make sure your editor inherits the same version.

2. Load the Chrome extension

1. Download or clone this repository.
2. Open chrome://extensions and turn on Developer mode.
3. Choose Load unpacked and select the chrome-extension directory.

That is the whole setup. There is no second server to start — the MCP server runs the connector itself.

3. Use it

Open Chrome DevTools (F12) on the page you want to inspect. Capture begins as soon as DevTools is open; the BrowserTools panel is only for settings and status. Then ask your agent something like "check the console for errors" or "run an accessibility audit on this page".

Not working? Run npx @agentdeskai/browser-tools-mcp --doctor, which reports exactly which piece is missing.

To watch capture happen live — useful when checking a fresh install — start the connector with --verbose:

npx @agentdeskai/browser-tools-server --verbose

· console error tab 42 Uncaught TypeError: total is not a function
· network 500 POST tab 42 https://myapp.local/api/pay (1310ms)

Without it the connector only reports connect and disconnect, so a working setup and a silent one look the same.

Tools

Three prompts ship alongside them — debuggerMode, auditMode and nextjsSeoAudit — giving your agent a systematic workflow instead of a wall of static text in every tool listing.

All tools declare MCP output schemas, so clients receive structured data rather than prose they have to parse, and read-only tools are annotated as such so clients can auto-approve them safely.

Several tabs at once

Every tab with DevTools open is tracked separately. Telemetry is attributed to the tab that produced it, and retention is per tab, so a chatty page cannot push out the history of the one you care about.

Tools act on the current tab — the one you most recently opened DevTools on. A tab whose connection drops and comes back does not steal that position, which is what used to make screenshots capture the wrong page. Every result reports the tabId and url it came from, plus otherTabs, so a wrong-tab answer is visible rather than silent. To target a specific tab, call listBrowserTabs and pass its tabId to any tool; pass allTabs: true to read across every tab.

Large data stays out of the context window

Whole-history payloads are exposed as MCP resources rather than inlined, and tools link to them with resource_link so your agent fetches them only when it decides to:

Log tools attach a link when a read had to be cut short; network reads always offer the HAR; screenshots always link to the stored image, which is the only way to see one too large to inline. The last 20 full Lighthouse reports are kept under audits/ in the screenshot directory.

Keeping responses small

Log payloads are the usual cause of a blown context window. Every read tool takes limit and offset, and the log tools take keyword filters:

getConsoleErrors({ keywords: ["hydration"], limit: 20 })
getNetworkLogs({ urlKeywords: ["/api/"], bodyKeywords: ["quota"], limit: 10 })

Results are returned newest-first and always report total alongside returned, so an agent knows when it is only seeing part of the picture.

Privacy and security

This tool captures whatever your browser sees, so it treats that data carefully:

• Loopback only. The connector binds 127.0.0.1 and refuses non-loopback addresses. In 1.x it bound 0.0.0.0, reachable by anyone on your network.
• The extension never leaves loopback. 1.x scanned private network ranges and adopted whichever host answered with a known string — meaning anyone on shared Wi-Fi could receive your logs and screenshots. That scan is gone.
• Authenticated. The HTTP API requires a per-run token. The WebSocket accepts browser-extension origins only, so a web page you visit cannot impersonate the extension.
• Credentials are scrubbed on the way in: Authorization, Cookie and similar headers, plus JWTs, cloud keys and vendor tokens found anywhere in captured strings, become [REDACTED].
• Headers are off by default, per direction, and storage values are withheld unless explicitly requested.
• Cookie access is an optional permission you grant from the panel, not something the extension holds by default.
• Audits stay off your internal network. An audit accepts a url, and the caller may be a model reading untrusted pages. Link-local and cloud metadata addresses are never audited. Pages on localhost or a private network are audited only while open in a connected tab (or allowlisted with --audit-allow-origin), and every request the audit browser makes — redirects and subresources included — is checked again as it happens. A local page may load only the local origins that are open in a tab or allowlisted, so if your app calls an API on another local port, open that too or allowlist it. Audits connect directly, not through a system proxy. Behind a fake-IP proxy (Clash, sing-box, Surge), which hides where names really go, audits are refused unless you pass --audit-allow-private-network; see SECURITY.md.

Report vulnerabilities per SECURITY.md.

Configuration

Flags, or the matching BROWSER_TOOLS_* environment variables:

To share one browser session between several MCP clients, start the connector once with npx @agentdeskai/browser-tools-server and every client will attach to it automatically.

Known limits

• Network capture starts when DevTools opens. Requests that finished before then are not recorded — reload the page to capture a full page load.
• Screenshots are held to a byte budget (screenshotMaxBytes, 3 MB by default). A capture that would exceed it is re-encoded as JPEG and, if still too large, downscaled. A viewport capture of dense content on a high-DPI display can otherwise run past 13 MB, which is more than a model's context can take and past the read buffer newer MCP stdio transports enforce. If an image still cannot fit, it is written to disk and the tool returns the path instead of inlining it.
• Console capture defaults to the DevTools protocol, which makes Chrome show a "started debugging this browser" banner. Switch the panel's capture mode to Wrap page console to avoid it.
• Firefox is not verified. The extension is written cross-browser — a browser/chrome shim, browser_specific_settings, and a capture mode that does not need chrome.debugger — but it has never been loaded in Firefox, and nothing in the test suite covers it. Screenshots in particular go through the DevTools protocol and will not work there. Treat Firefox as unsupported until someone has actually run it; a report either way is welcome.
• Audits launch a separate browser and take up to a minute. Any Chromium-based browser works — Chrome, Chromium, Brave, Edge, Vivaldi, Opera or Arc — and --doctor reports which one will be used. Set CHROME_PATH to override. Arc is supported on a best-effort basis and has not been verified headless.

Development

npm install
npm run build
npm test           # unit + integration, no browser required
npm run test:e2e   # real Chromium with the extension loaded

npm test runs in seconds. The end-to-end suite launches a headed Chromium with the extension installed, drives fixture pages, and asserts the whole capture path — run npx playwright install chromium first.

License

MIT

内容由编辑整理，部分信息来自 GitHub 与项目官网，版权归原作者所有。发现错误？报告问题