twin-browser-mcp

The Twin Browser MCP server. Drive real browsers for your AI agents — run intent-level goals, extract structured data, compile a workflow once and replay it at near-zero cost — from any MCP client (Claude Desktop, Cursor, Claude Code, Cline, …). No local browser, no engine to run: it talks to the hosted Twin Browser API under your key.

Setup

You need one thing: a Twin Browser API key (ab_live_…) from your dashboard.

Claude Desktop / Cursor — add to your MCP config:

{
  "mcpServers": {
    "twin-browser": {
      "command": "npx",
      "args": ["-y", "twin-browser-mcp"],
      "env": { "TWIN_API_KEY": "ab_live_…" }
    }
  }
}

Claude Code / Cline:

claude mcp add twin-browser --env TWIN_API_KEY=ab_live_… -- npx -y twin-browser-mcp

That's it. WEB_BASE_URL defaults to https://twin-browser.com; every tool routes through the hosted API under your key.

Tools

30 tools, including:

• run_goal — run an intent-level goal against a URL until a success condition is met
• extract — cheap structured extraction (fields or a JSON schema)
• compile_skill / run_skill — compile a workflow once, replay deterministically with no LLM
• dispatch — semantic-cache a request to a compiled skill
• observe_page, screenshot, search, map_site, submit_crawl
• connect_account — get a one-time link so a user signs into a gated site by hand
• submit_verification, control_run — drive a paused run (2FA codes, direct input)
• monitors, ETL, and more

Run the server once to see the full list printed to stderr, or read the docs.

Operator mode (self-hosted engine)

If you run your own Twin Browser execution engine, set EXEC_URL + EXEC_SHARED_SECRET instead of (or alongside) TWIN_API_KEY and the server talks to your engine directly. Most users don't need this — the hosted default is one key.

Links

• Docs: https://twin-browser.com/docs/mcp
• Dashboard / API keys: https://twin-browser.com/dashboard/keys

MIT