About Tailnet MCP bridge
Tailnet MCP bridge is a stdio MCP server for hosted agents and sandboxed clients that need to reach an MCP server inside a Tailscale tailnet. It embeds a Tailscale node and forwards requests to a Streamable HTTP endpoint without exposing the server to the public internet.
Key features
- 01Bridge stdio to Streamable HTTP
- 02Join a tailnet as a tagged device
- 03Keep the MCP server off the public internet
- 04Recover sessions after server restarts
- 05Retry only requests that never arrived
Use cases
- 01Connect a hosted agent sandbox to a tailnet server
- 02Reach an MCP server published through tailscale serve
- 03Maintain MCP sessions across server restarts
MCP support
Acts as a stdio MCP server and forwards requests to a Streamable HTTP MCP server inside the tailnet. It bridges the client's MCP connection rather than exposing tools of its own.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
docker run -i --rm -e TS_AUTHKEY ghcr.io/zedong-peng/tailnet-mcp:v0.3.0TS_AUTHKEYSecretTagged, pre-authorized auth key; only needed the first timeLocal servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
ghcr.io/zedong-peng/tailnet-mcp:v0.3.0stdioTools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project READMEtailnet-mcp
A stdio MCP server that forwards everything to a Streamable HTTP MCP server inside your tailnet.
Hosted agents (Grok, cloud coding agents, sandboxed assistants) usually take MCP servers two ways. By URL: the vendor's servers connect, so your server has to be on the public internet. By command: the agent starts a process on its own computer, but that computer is not in your tailnet and you can't run tailscaled there (no root, no TUN). tailnet-mcp is the command: it embeds a Tailscale node (tsnet), joins your tailnet as a tagged device and bridges stdio to the server. Your server stays behind tailscale serve, never on the internet.
agent ──stdio──▶ tailnet-mcp (tag:agent) ══tailnet══▶ tailscale serve ──▶ MCP server on 127.0.0.1
Install
Download a static binary for linux or darwin, amd64 or arm64, from Releases and check it against SHA256SUMS. Or build it yourself:
CGO_ENABLED=0 go install github.com/zedong-peng/tailnet-mcp@latest
Or run the image ghcr.io/zedong-peng/tailnet-mcp with a volume on /home/nonroot/.local/state/tailnet-mcp. It is also listed in the MCP Registry as io.github.zedong-peng/tailnet-mcp.
Use
-
Publish the server to the tailnet only, on a path nobody can guess:
tailscale serve --bg --https=8443 --set-path=/$(openssl rand -hex 16) http://127.0.0.1:8080/mcp -
Mint a one-time, pre-authorized auth key with a tag for the agent (admin console, or the API), and log in once from the agent's computer:
TS_AUTHKEY=tskey-auth-... tailnet-mcp -hostname my-agent -url https://server.tailnet.ts.net:8443/<path> -check-checkconnects, prints the server's tools on stderr and exits. The node key is kept in-state(default~/.local/state/tailnet-mcp/<hostname>); keep it on persistent storage. Tagged nodes don't expire, so the key is not needed again. -
Register the same command, without the key, as the agent's MCP server:
tailnet-mcp -hostname my-agent -url https://server.tailnet.ts.net:8443/<path>
Each process is its own tailnet node. For a second server, run a second process with a different -hostname.
Lock it down
The agent reads untrusted input (web pages, mail), so assume it can be talked into anything its tools allow. Three layers, none of which rely on the agent behaving:
-
Network. Give the tag one grant to the ports it needs and nothing else:
"grants": [ {"src": ["tag:my-agent"], "dst": ["server"], "ip": ["tcp:8443"]}, ], "tests": [ {"src": "tag:my-agent", "accept": ["server:8443"], "deny": ["server:22", "server:443"]}, ],Grants that allow everything (
"src": ["autogroup:member"], "dst": ["*"]) don't cover tagged devices, so that one grant is all the tag gets. -
Path.
tailscale serve --set-pathanswers only on the secret path; any other path on the port is a 404. Other devices in the tailnet can't use the server by guessing the port. -
Server. Enforce permissions where the data lives, not in the agent: a mail server that can only save drafts, a password vault account that is a view-only member of one shared collection. examples/private-tools sets up both.
Behaviour
- Before the first request it pings the server's node until it answers. Right after a restart the server may still hold a stale path to the node, and the first dial would time out.
- Only requests that never reached the server are retried, so a tool call is never sent twice.
- When the server restarts and forgets the session (404, or a 400 about the session), the bridge replays
initializeand resends the request. The client doesn't notice. - A failed request comes back as a JSON-RPC error with the reason, so the client never waits for an answer that won't come.
- On EOF it ends the session with
DELETE. Logs go to stderr; stdout carries only MCP frames.-vadds tsnet's own logs.
Not supported: the optional GET stream that lets a server send messages outside a response. Servers answer each request in its own response, as almost all tool servers do.
Compared with
mcp-remoteand other stdio-to-HTTP bridges need a network path to the server. In an agent's sandbox that means a public URL.tailscale upin the sandbox needs root and a TUN device. If someone logs in as a user instead of with a tagged key, the sandbox gets that user's access to the whole tailnet.- Tailscale Aperture is a gateway on the server side: it aggregates HTTP MCP servers and grants tools per user. It doesn't get a sandboxed agent into the tailnet, so the two fit together.
Build and test
go test -race ./...
CGO_ENABLED=0 go build -trimpath -ldflags="-s -w" .
Tagging v* publishes the four binaries, the image and the registry entry.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.