Browser MCP
https://github.com/browsermcp/mcp- Category
- Developer Tools
- Rank
- No. 1020Tools index
- Listed in
- #9 Control a browser
- Pricing
- Open Source
- Platform
- desktop · extension
- Type
- TOOL
- Builder
- browsermcp
- GitHub
- 7.0k stars
- Added
- Aug 12, 2026
About
An MCP server paired with a Chrome extension that lets AI applications control a user's actual browser locally. It preserves logged-in sessions and uses the real browser fingerprint instead of launching a fresh automated instance.
What it does
Browser MCP gives an AI client commands for navigation, clicking, typing, selection, keyboard input, waiting, screenshots, console logs, and accessibility snapshots. The server communicates over standard input and output, while browser actions travel through a WebSocket connection to an attached Chrome tab.
Why it's ranked here
The tool covers a practical browser interaction loop and returns page state after most actions, which helps an AI continue reliably. The public repository is less convincing as a reusable codebase because its own documentation says required monorepo packages prevent standalone builds.
What's good
Accessibility snapshots include the current URL, title, and structured page content. Navigation and interaction commands often return a fresh snapshot immediately. Inputs are validated against schemas, tool errors become structured MCP errors, and screenshots use native image responses instead of text encoding.
Tradeoffs
A user must manually connect a browser tab through the extension before commands work. The server exposes no resources, and one implemented drag action is not registered among its available tools. Startup forcibly kills the process occupying its configured WebSocket port. The published source also depends on unavailable workspace packages.
How to use it well
Use it when an AI assistant needs to operate sites where your existing Chrome state matters, then inspect accessibility snapshots or screenshots between actions. It fits interactive browsing, form work, and logged-in workflows. It does not provide an isolated fresh-browser environment, and this repository alone is not suitable for independent source builds.
Technical notes+
src/index.ts starts an MCP stdio server and registers navigation, history, snapshot, click, hover, typing, option selection, keypress, wait, console-log, and screenshot tools; src/tools/snapshot.ts also defines drag, but the entry point omits it. src/context.ts routes requests through the extension WebSocket with a 30-second default timeout. src/ws.ts calls killProcessOnPort, whose implementation in src/utils/port.ts force-kills the configured port owner before binding. src/server.ts replaces server.close with a closure that calls server.close again, creating an apparent recursive shutdown defect. package.json references several workspace:* development dependencies, matching the standalone-build warning in README.md.
Observed
- License
- Apache License 2.0
- Primary language
- TypeScript
- Packaging
- ES module npm package with a command-line executable and compiled dist-only publication surface
- Interfaces
- MCP over standard input and output, plus WebSocket communication with the browser extension
- Browser support
- Chrome extension
- MCP capabilities
- Tools and resources are declared, but the configured resource list is empty
- Build structure
- The repository cannot build independently because it relies on workspace packages from its development monorepo
Read from README.md, package.json, src/index.ts, src/server.ts, src/ws.ts, src/context.ts, src/utils/log.ts, src/tools/tool.ts, src/utils/port.ts, src/tools/common.ts, src/tools/custom.ts, src/tools/snapshot.ts, src/resources/resource.ts, src/utils/aria-snapshot.ts, LICENSE.
Tags
Tech Stack
Comments (0)
No comments yet
Indexed by a proprietary survey. Corrections welcome.