browser-mcp

A minimal, reliable browser control MCP server for LLM agents.

Built to fix the most common pain points with Playwright MCP: too many tools, zombie processes, memory leaks, and fragile setup.

Features

• 9 focused tools instead of 25+ — reduces agent confusion
• Auto-cleanup on exit — no zombie Chrome processes
• Auto-reset if memory exceeds 500MB — no runaway leaks
• Fuzzy matching — click and type by visible text or label, not just CSS selectors
• One-command setup

Install

npx browser-mcp install

Or manually:

git clone https://github.com/obsoul/browser-mcp
cd browser-mcp
npm install
npx playwright install chromium
npm run build

Add to Claude Code

Add this to your ~/.claude.json under mcpServers:

"browser": {
  "command": "node",
  "args": ["C:/path/to/browser-mcp/dist/index.js"],
  "type": "stdio"
}

Or if published to npm:

"browser": {
  "command": "npx",
  "args": ["browser-mcp"],
  "type": "stdio"
}

Tools

| Tool | Description | |---|---| | navigate | Go to a URL | | snapshot | Get page title, headings, and interactive elements | | click | Click by CSS selector, text, or aria-label | | type | Type into a field by selector, placeholder, or label | | extract | Extract tables, links, or text from the page | | screenshot | Capture the current page | | wait | Wait for a selector, text, or N milliseconds | | reset | Kill browser and start a fresh clean session | | status | Get current URL, title, and connection state |

Example usage in Claude Code

navigate to https://news.ycombinator.com
snapshot
extract links

Why not just use Playwright MCP?

| Problem | This server | |---|---| | 25+ tools confuse agents | 9 tools only | | Zombie Chrome processes | Cleanup on SIGINT/SIGTERM/exit | | 24GB memory leaks | Auto-reset at 500MB | | Profile lock errors | Single managed context, no user profile | | Opaque failures | Returns what it actually matched |

License

MIT