FR: Feature request: Official MCP server for Tailscale administration

### What are you trying to do?

# Feature request: Official MCP server for Tailscale administration

> [!NOTE]
> I used an AI assistant to help draft and polish this feature request so it would be clear and easy to read. I fully reviewed the text, agree with it, and am submitting it as my own request.

I’d like an official Tailscale MCP server that exposes scoped Tailscale administration tools to MCP-compatible chatbots and agents, including both local and cloud-hosted assistants.

## Use case

I want to be able to ask a chatbot to inspect and help administer my tailnet through the Tailscale API, with explicit confirmation before any write operation. For example:

- Read current ACL/policy file
- Validate proposed ACL/policy changes
- Preview policy effects before applying
- Apply policy changes after user confirmation
- List devices, users, tags, routes, DNS settings, and posture data
- Help diagnose access issues by correlating ACLs, device state, tags, routes, and DNS

This would be especially useful for homelab users and small teams where the policy file can become complex, but where granting a general-purpose assistant broad credentials is risky.

## Desired properties

- Officially maintained by Tailscale
- Supports OAuth clients / trust credentials rather than long-lived full API keys
- Fine-grained scopes for read-only, policy-write, device-admin, DNS-admin, route-admin, etc.
- Read-only mode by default
- Write operations require explicit confirmation from the MCP client/user
- Policy updates should use validate/preview endpoints before apply
- Clear audit logging of tool calls and applied changes
- Safe transport options for local use, Tailscale Serve, or Aperture
- Compatible with common MCP clients such as Claude Desktop, Cursor, Codex, and local LLM frontends

There are already several community Tailscale MCP servers, which suggests demand exists, but an official server would be easier to trust for security-sensitive administration.

This seems aligned with Tailscale’s recent work around Aperture MCP proxying and identity-aware AI tooling, but focused specifically on administering Tailscale itself through well-scoped MCP tools.

This would be a big ask I'm sure, but it would realy be cool and helpful (if written with gaurdrails)

### How should we solve this?

_No response_

### What is the impact of not solving this?

_No response_

### Anything else?

_No response_