MCP Server

The meshbrow-mcp server exposes Meshbrow’s stealth browser capabilities as MCP tools. Compatible with Claude Desktop, VS Code Copilot, Cursor, and any MCP-compatible client.

Installation

Homebrew (macOS/Linux)
Go Install
Binary
Docker

brew tap meshbrow-dev/tap
brew install meshbrow-mcp

go install github.com/meshbrow-dev/meshbrow-mcp@latest

curl -fsSL https://github.com/meshbrow-dev/meshbrow-mcp/releases/latest/download/meshbrow-mcp-linux-amd64 \
  -o /usr/local/bin/meshbrow-mcp
chmod +x /usr/local/bin/meshbrow-mcp

docker pull ghcr.io/meshbrow-dev/meshbrow-mcp:latest

Verify the installation:

meshbrow-mcp --version
# meshbrow-mcp v0.1.0

Configuration

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "meshbrow": {
      "command": "meshbrow-mcp",
      "args": ["--mode", "stdio"],
      "env": {
        "MESHBROW_API_KEY": "mb_live_your_key_here"
      }
    }
  }
}

VS Code (GitHub Copilot)

Add to .vscode/mcp.json in your project:

{
  "servers": {
    "meshbrow": {
      "command": "meshbrow-mcp",
      "args": ["--mode", "stdio"],
      "env": {
        "MESHBROW_API_KEY": "mb_live_your_key_here"
      }
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "meshbrow": {
      "command": "meshbrow-mcp",
      "args": ["--mode", "stdio"],
      "env": {
        "MESHBROW_API_KEY": "mb_live_your_key_here"
      }
    }
  }
}

Standalone (WebSocket)

Run the MCP server over WebSocket for remote/shared access:

meshbrow-mcp --mode ws --port 9090 --api-key mb_live_...

CLI Flags

Flag	Env Variable	Default	Description
`--mode`	`MESHBROW_MCP_MODE`	`stdio`	Transport: `stdio` or `ws`
`--port`	`MESHBROW_MCP_PORT`	`9090`	WebSocket listen port
`--api-url`	`MESHBROW_API_URL`	`https://api.meshbrow.dev`	API endpoint
`--api-key`	`MESHBROW_API_KEY`	(required)	Your API key
`--log-level`	`MESHBROW_LOG_LEVEL`	`info`	Log level

Available Tools

Session Management

Tool	Description
`session_create`	Launch a stealth browser session with proxy and fingerprint
`session_list`	List all active sessions
`session_get`	Get session details and CDP endpoint
`session_destroy`	Destroy a session (optionally save profile)

Browser Actions

Tool	Description
`browser_navigate`	Navigate to a URL
`browser_screenshot`	Take a screenshot (base64 PNG)
`browser_click`	Click an element by CSS selector
`browser_type`	Type text into an input field
`browser_extract`	Extract text/data from the page
`browser_execute`	Execute arbitrary JavaScript
`browser_wait`	Wait for an element to appear
`browser_scroll`	Scroll the page

Profiles

Tool	Description
`profile_create`	Create a persistent browser profile
`profile_list`	List saved profiles
`profile_get`	Get profile details
`profile_delete`	Delete a profile

Fleet

Tool	Description
`fleet_create`	Launch multiple sessions in parallel
`fleet_status`	Check fleet session statuses
`fleet_destroy`	Destroy all fleet sessions

Example: AI Agent Web Research

User: "Find the current price of Bitcoin on CoinGecko"

Agent calls: session_create({ stealth: "max", proxy_country: "US" })
→ Returns: { id: "sess_abc123", cdp_endpoint: "wss://..." }

Agent calls: browser_navigate({ session_id: "sess_abc123", url: "https://coingecko.com/en/coins/bitcoin" })
→ Returns: { status: "loaded" }

Agent calls: browser_extract({ session_id: "sess_abc123", selector: "[data-target='price.price']" })
→ Returns: { result: "$67,432.12" }

Agent calls: session_destroy({ session_id: "sess_abc123" })
→ Returns: { status: "destroyed" }

Agent responds: "The current Bitcoin price on CoinGecko is $67,432.12"

Security

The MCP server enforces:

API key required — all tool calls require a valid MESHBROW_API_KEY
Session isolation — each session runs in its own network namespace
No credential leakage — proxy credentials and fingerprint internals are never exposed in tool responses
Structured logging — all operations are logged to stderr (never stdout, which is reserved for MCP protocol)

Best Practices

Use profiles for authenticated workflows

Store login cookies in profiles so agents don’t need to re-authenticate each time.

Always destroy sessions

Include cleanup in your agent’s error handling. Leaked sessions cost money and resources.

Prefer extract over screenshots

Text extraction is faster and uses fewer tokens than screenshot-based approaches.

​MCP Server

​Installation

​Configuration

​Claude Desktop

​VS Code (GitHub Copilot)

​Cursor

​Standalone (WebSocket)

​CLI Flags

​Available Tools

​Session Management

​Browser Actions

​Profiles

​Fleet

​Example: AI Agent Web Research

​Security

​Best Practices