Tailnet MCP Server

Why I built this: I built this to learn Tailscale - this was my first time using it. From past experience debugging network issues, I thought having an MCP to feed network state directly into an LLM could make diagnosing connectivity problems much easier than context-switching between CLI output and chat windows.

A Model Context Protocol (MCP) server that exposes your Tailscale tailnet as a queryable resource for LLMs. Query devices, check connectivity, and get network status directly from your AI assistant.

Prerequisites

• Docker

• A Tailscale account with at least one device on your tailnet

Related MCP server: mcp-tailscale

Setup & Test

1. Clone & enter the repo:

   git clone https://github.com/AlexSKorn/TailscaleMCP.git
   cd TailscaleMCP

2. Get an Auth Key: Go to your Tailscale Admin Console, click "Generate auth key", and copy the key (starts with tskey-auth-).

3. Configure & start:

   make setup                          # creates .env from template
   # edit .env and paste your TS_AUTHKEY
   make docker-up                      # start containers
   make docker-test                    # should print "Connection test: OK"

4. Try it interactively with the MCP Inspector:

   npx @modelcontextprotocol/inspector docker compose exec -T tailnet-mcp python -m tailnet_mcp.server

   Open the URL shown in your terminal (usually http://localhost:3000), click Connect and then "List Tools", then pick any tool and click "Run".

Available Tools

Cleanup

make docker-down    # stop containers (keeps state)
make docker-clean   # stop and remove all data

Architecture

MCP Client (stdio) ←→ tailnet-mcp (FastMCP) ←→ tailscale (Unix socket)
                                                    ↓
                                            /localapi/v0/*

1. MCP client sends tool calls via stdio

2. Server translates them to Tailscale Local API requests over a Unix socket

3. Tailscale daemon responds with network state

4. Server formats responses for human readability

Key design decisions:

• Read-only — the server cannot modify Tailscale settings, add/remove devices, or change ACLs

• Local API only — all calls stay on the local machine (no external network calls, no stored credentials)

• Async throughout — httpx for socket I/O, subprocess for pings

Project Structure

src/tailnet_mcp/
├── server.py        # MCP server & tool definitions
├── tailscale.py     # Tailscale local API client (async httpx over Unix socket)
├── models.py        # Pydantic data models
tests/
└── test_server.py   # pytest + pytest-asyncio tests

Development

pip install -e ".[dev]"    # install with dev dependencies
make test                  # run tests
make check-all             # lint + format + type-check + tests

License

MIT