Server Agent Browser

Status: Testing — Functional but still being validated. Expect rough edges.

Let LLMs drive a real browser. This MCP server wraps agent-browser (by Vercel Labs) so any MCP-compatible client — Claude Desktop, Claude Code, or others — can navigate pages, fill forms, click buttons, take screenshots, and more.

Prerequisites

Install the agent-browser CLI:

cargo install agent-browser
agent-browser install

Build

cargo build --release

Binary: target/release/mcp-server-agent-browser

Configuration

Add to your MCP client config (example for Claude Desktop / Claude Code):

{
  "mcpServers": {
    "agent-browser": {
      "command": "/path/to/mcp-server-agent-browser"
    }
  }
}

The server communicates over stdio using JSON-RPC 2.0 — it works with any MCP client.

Environment Variable	Description	Default
AGENT_BROWSER_PATH	Path to the agent-browser binary	agent-browser (via $PATH)

What it can do

The server exposes 35+ tools covering:

• Navigation — open URLs, go back/forward, reload
• Interaction — click, type, fill, press keys, hover, scroll, drag, select, check/uncheck, upload, download
• Reading — get page text, HTML, attributes, URL, title
• Accessibility — browser_snapshot returns an accessibility tree with @ref identifiers that can be used as selectors in subsequent calls
• State — check if elements are visible, enabled, or checked
• Capture — screenshots (with optional annotation for vision models) and PDF export
• Sessions — create isolated browser sessions with independent cookies/storage/viewport
• Cookies — get, set, clear
• JavaScript — evaluate arbitrary JS in the page context
• DevTools — read console logs, inspect network requests, connect via CDP

Each tool is self-documented with parameter schemas — your MCP client will discover them automatically.

How it works

A typical AI-driven interaction follows this pattern:

1. browser_navigate — open a page
2. browser_snapshot — get the accessibility tree to understand the page structure
3. browser_click / browser_fill — interact using @ref selectors from the snapshot
4. browser_screenshot — visually verify the result

The agent decides which tools to call and in what order. You just describe what you want done.

Sessions

All tools accept an optional session_id. Create isolated sessions with browser_new_session and pass the returned ID to subsequent calls. Each session has its own cookies, storage, and viewport. Without a session ID, commands use the default session.

Timeouts

Commands time out after 60 seconds by default. The browser_wait tool adjusts its timeout automatically based on the wait duration you specify.

Limitations

• Requires the Rust toolchain to build, and Node.js for the agent-browser daemon
• Chromium only — agent-browser uses Playwright under the hood
• Headless by default — no visible browser window
• The 60-second default timeout may not be enough for slow pages or complex interactions
• No built-in authentication persistence — cookies reset between server restarts unless managed via the cookie tools
• File upload/download paths are validated but restricted from system directories

License

MIT