DeskLink

Turn your local computer into a remotely accessible MCP endpoint for cloud AI agents.

DeskLink is a thin authenticated bridge. It starts the local STDIO MCP servers you already use (or the official filesystem server), discovers their tools, and exposes everything through a single Streamable HTTP MCP URL. Optionally tunnels that URL with ngrok so Claude, ChatGPT, Cursor, Manus, or any remote MCP client can reach your machine securely.

Package name on PyPI: desklink
CLI commands: desklink and devlink (alias)

What it does

1. Loads MCP servers from a standard mcp.json (or injects the official @modelcontextprotocol/server-filesystem if none are defined).

2. Starts a local Streamable HTTP server (default http://127.0.0.1:8765/mcp).

3. Proxies every discovered tool under a namespaced name (servername__toolname).

4. Protects the endpoint with a device bearer token and/or OAuth 2.1 (Authorization Code + PKCE).

5. Optionally opens an ngrok tunnel so the same endpoint is reachable from the public internet.

6. Gives you a simple web dashboard + consent screen for OAuth apps.

You keep full control: tools stay on your machine, approvals can prompt you, and tokens never leave your machine unless you choose.

Related MCP server: Quota

Quick start

pip install desklink

# Optional but recommended for public tunnels
export NGROK_AUTHTOKEN=your_ngrok_token

desklink start

On first run DeskLink will:

• Create ~/.desklink/token (device bearer token)

• Start the official filesystem MCP server scoped to your home directory (unless you provide an mcp.json)

• Print the local and public MCP URLs plus the auth token

Copy the public URL + token into any MCP-compatible client.

Useful flags

desklink start \
  --mcp-config ./mcp.json \
  --config ./desklink.config.json \
  --dir ~/Documents --dir ~/projects \
  --port 8765 \
  --tunnel ngrok          # or "none" for local-only
  --token my-static-token # optional override
  -v                      # verbose logging

Configuration

DeskLink looks for two files (first match wins):

Example mcp.json

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Documents", "/Users/you/projects"]
    },
    "my-custom-server": {
      "command": "python",
      "args": ["-m", "my_mcp_server"],
      "env": { "API_KEY": "..." }
    }
  }
}

Any STDIO MCP server works. Tools appear as filesystem__read_file, my-custom-server__do_something, etc.

Example desklink.config.json

{
  "logging": { "level": "info" },
  "tracing": {
    "enabled": true,
    "level": "basic",
    "redact_sensitive": true
  },
  "auth": {
    "enabled": true
  },
  "approvals": {
    "default": "auto",
    "tools": {
      "filesystem__write_file": "prompt",
      "filesystem__create_directory": "prompt"
    }
  },
  "server": {
    "host": "127.0.0.1",
    "port": 8765,
    "mcp_path": "/mcp"
  },
  "tunnel": {
    "provider": "ngrok"
  }
}

Approvals

• "auto" – run immediately

• "prompt" – ask in the terminal before the tool executes

Tracing levels

• none / basic / full (full logs args + truncated results; sensitive keys are redacted by default)

Environment variables use the prefix DESKLINK_ (e.g. DESKLINK_SERVER__PORT=9000). NGROK_AUTHTOKEN is also picked up automatically.

Authentication

1. Device bearer token (always available)

Stored at ~/.desklink/token. Sent as:

Authorization: Bearer <token>

Use this for simple scripts or clients that just need a static secret.

2. OAuth 2.1 (recommended for cloud agents)

DeskLink acts as a full OAuth 2.1 authorization server so agents never see your device token.

Create an app (CLI)

desklink create oauth-app claude
desklink create oauth-app chatgpt
desklink create oauth-app manus
desklink create oauth-app cursor
desklink create oauth-app custom --name "My Agent" --public

Presets come with sensible default scopes. Confidential clients get a Client Secret; public clients rely on PKCE.

Endpoints (when desklink start is running)

When using a tunnel the same paths work on the public URL.

Manage apps

desklink oauth list
desklink oauth revoke <client_id>

Or open /dashboard in a browser – create apps, view Client IDs, revoke access, and see recent activity.

Scopes

All OAuth state lives as JSON under ~/.desklink/oauth/ (no database).

Customer usage flow

1. Install & start DeskLink on the machine whose tools/files you want to expose.

2. (Optional) Create an OAuth app for the agent you will connect.

3. Give the agent the public MCP URL + either the device token or the OAuth Client ID (and secret if confidential).

4. The agent discovers tools via normal MCP tools/list and calls them. DeskLink proxies the call to the local STDIO server, applies your approval policy, and returns the result.

5. Stop with Ctrl+C. The tunnel and local servers shut down cleanly.

Extensibility

• Any STDIO MCP server – just add it to mcp.json. No code changes required.

• Approvals & tracing – configure per-tool policies and log levels without touching the bridge.

• Tunnel providers – currently ngrok and none. The interface is abstract; additional providers (Cloudflare Tunnel, etc.) can be added later.

• Auth – static bearer + full OAuth 2.1 (PKCE, refresh tokens, dynamic registration, consent screen). Device identity stays separate from OAuth application credentials.

• Programmatic use – import desklink and drive DeskLinkServer / McpBridge from your own Python process if you need tighter integration.

Development

git clone https://github.com/sam-14uel/desklink-py.git
cd desklink-py
pip install -e ".[dev]"
pytest

License

MIT © Samuel Obinna Chimdi