OpenClaw MCP Bridge

A high-performance bridge that exposes OpenClaw capabilities as a Model Context Protocol (MCP) server.

This bridge allows remote AI agents (like OpenCode, Cursor, or Windsurf running on remote servers) to securely delegate local tasks—such as browser automation and shell execution—to an OpenClaw Gateway running on your local machine (e.g., MacBook) via Tailscale or any network connection.

Why this exists

When working on powerful remote servers (GPU clusters, cloud instances, etc.), AI agents are "blind" to your local environment. This bridge provides the "eyes and hands" by connecting the remote agent to your local OpenClaw instance.

• Local Browser Control: Navigate, screenshot, click, and scrape pages on your local machine.

• Local Shell Access: Execute commands on your local OS (say, open, pbcopy, hardware control).

• Secure Authentication: API key via HTTP headers + OpenClaw Ed25519 challenge-response signing.

• Hybrid Routing: Intelligently routes commands between the OpenClaw Gateway and connected Nodes.

Related MCP server: mcp-devtools

🎯 Use Cases

1. Remote Development with Local Eyes

Run heavy AI models or IDE backends on a powerful cloud server while letting the AI see and control your local browser for previewing changes.

2. "Job Done" Physical Notification (HomePod)

Combine this bridge with homepod-announcer-skill to let your remote AI agent speak to you when a long task completes.

• Remote Agent: "Training finished. Sending notification to local HomePod."

• Local Bridge: Receives request -> Forwards to local HomePod Skill.

• HomePod: "Sir, your model training is complete with 98% accuracy."

Architecture

graph LR
    subgraph Remote_Server
        A[AI Agent / OpenCode] -- "SSE (MCP)" --> B[openclaw-bridge-remote]
    end

    subgraph Local_Machine
        B -- "Tailscale/Network" --> C[OpenClaw Gateway]
        C -- "WebSocket" --> D[Local Browser]
        C -- "WebSocket" --> E[OpenClaw Node]
        E -- "Exec" --> F[Local OS Shell]
    end

Quick Start

1. Install (on your local machine, e.g. MacBook)

curl -fsSL https://raw.githubusercontent.com/lucas-jo/openclaw-bridge-remote/main/install.sh | bash

The installer will:

• Clone the repo to ~/openclaw-bridge-remote

• Install dependencies via Bun

• Auto-detect your OPENCLAW_GATEWAY_TOKEN

• Generate a random BRIDGE_API_KEY

• Create a .env file with all settings

2. Verify the bridge is running

# From your remote machine (replace with your local machine's IP)
curl http://<your-local-ip>:3100/health
# Expected: {"status":"ok","gateway":true,"version":"0.1.0"}

3. Get the values you need for MCP config

You need two pieces of information from the .env file on your local machine:

# On your local machine (where the bridge is running)
cat ~/openclaw-bridge-remote/.env

Note down:

• BRIDGE_API_KEY — The secret key for authenticating MCP clients

• Your local machine's IP — Use tailscale ip -4 (Tailscale) or your LAN IP

4. Configure your AI agent (on the remote machine)

Important: Use headers with x-api-key for authentication, NOT query parameters in the URL. The MCP SDK's SSE transport drops query parameters from the URL when making POST requests, which causes authentication failures.

OpenCode (~/.config/opencode/config.json):

{
  "mcp": {
    "openclaw-remote": {
      "type": "remote",
      "url": "http://<your-local-ip>:3100/sse",
      "headers": {
        "x-api-key": "<your-BRIDGE_API_KEY>"
      },
      "enabled": true
    }
  }
}

Cursor / Windsurf / Claude Code (MCP config):

{
  "mcpServers": {
    "openclaw-remote": {
      "type": "remote",
      "url": "http://<your-local-ip>:3100/sse",
      "headers": {
        "x-api-key": "<your-BRIDGE_API_KEY>"
      }
    }
  }
}

Local agents (stdio mode) — For agents running on the same machine as the bridge:

{
  "mcpServers": {
    "openclaw-local": {
      "command": "bun",
      "args": ["run", "start", "--transport=stdio"],
      "cwd": "~/openclaw-bridge-remote"
    }
  }
}

5. Test the connection

After configuring your agent, try calling any tool (e.g. openclaw_gateway_call with method: "node.list") to verify the bridge is working end-to-end.

Configuration Reference

The .env file controls the bridge behavior:

Running the Bridge

SSE Mode (for remote agents)

bun run start             # Default: SSE mode
bun run start --transport=sse  # Explicit

Stdio Mode (for local agents)

bun run start --transport=stdio

Available MCP Tools

Required: Browser Relay Configuration

If you run both the Gateway and Node Host (needed for openclaw_system_run), you must add this to ~/.openclaw/openclaw.json to prevent EADDRINUSE port conflicts on the browser relay:

{
  "gateway": {
    "nodes": {
      "browser": {
        "mode": "off"
      }
    }
  }
}

This tells the Gateway to handle browser requests directly instead of proxying them to the Node Host (which would try to bind the same relay port). See SETUP.md for details.

Documentation

• SETUP.md: Detailed architecture, protocol details, browser relay config, and troubleshooting.

• RECIPES.md: Connectivity options (Tailscale, SSH Tunneling, Cloudflare Tunnel).

Development

bun run dev       # Watch mode (auto-restart on changes)
bun run lint      # Check code style
bun run format    # Auto-fix code style
bun run typecheck # Check for type errors

Releasing

1. Update version in package.json.

2. Push a tag:

git tag v0.2.1
git push origin v0.2.1

The Release Workflow automatically runs checks and creates a GitHub Release.

License

MIT © Lucas Jo