@yawlabs/tailscale-mcp

Ask your agent questions about your tailnet and have it act on the answers. 97 admin-API tools + 6 optional local-CLI diagnostics + 1 always-on catalog tool + 4 resources spanning the Tailscale v2 API — devices, ACLs, DNS, keys and trust credentials, users, invites, webhooks, log streaming, posture, services, and organization tailnets. Backed by 1900+ unit tests and an opt-in live-tailnet integration suite.

Built and maintained by Yaw Labs.

One click adds this to your local Yaw MCP config so it's available in every Yaw Terminal session. Or install manually below.

What's the point if the API already exists?

You could curl the Tailscale API. The point isn't replacing curl — it's letting an agent compose multi-endpoint workflows in one turn without writing a script:

• "Which devices haven't checked in for 30 days and have key expiry disabled?" — lists devices, filters by lastSeen (online devices carry none), filters by keyExpiryDisabled, returns a table. Three endpoints, one question.
• "Someone broke DNS at 2am — who changed what in the last 24 hours?" — pulls the audit log, filters by DNS-related actors and endpoints, reads each change's before/after, summarizes in English.
• "Draft an ACL change that lets tag:mobile reach tag:dashboard but not tag:db, preserving my comments" — reads the current HuJSON, proposes a minimal diff, validates it against the API, returns the diff for you to apply.
• "Rotate every auth key older than 90 days and print the new ones" — iterates, creates new keys with matching tags, revokes the old ones.
• "Create an OAuth client for our CI pipeline scoped to devices:core:read and dns:read" — creates a trust credential via tailscale_create_key with keyType=client, returns the credentials once (save them immediately).

A curl can do each step. The agent composes them. That's where the lift is, and that's what the tool surface is designed for — every read endpoint is first-class so the agent can synthesize, and every write endpoint is tagged destructiveHint or idempotentHint so your MCP client can gate mutations the way you configured it.

If all you need is one endpoint in a CI job, use curl — we even have a CLI subcommand for the common ACL-from-git case. The MCP is for the interactive, exploratory, "I don't know what I need yet" work.

Why MCP vs. a skill or the tailscale CLI?

Reasonable question. Both have their place. Where this MCP is better:

• Broad admin API coverage. The tailscale CLI is scoped to the node it runs on. Admin concerns — ACLs, users, invites, webhooks, log streaming, posture integrations, auth keys, OAuth clients, and federated identities — live in the v2 HTTP API. You'd be shelling out to curl anyway.
• Typed tool surface, not string parsing. Every tool has a Zod-validated input schema and a structured response. No brittle tailscale status --json | jq pipelines that break when the schema evolves.
• Cross-client, and updates arrive as a version bump. An MCP server works in Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, and anything else that speaks MCP; a skill written to the Agent Skills standard travels between agents too, but it is prose you maintain. Version bumps ship through npx — nobody rewrites a skill's instructions when Tailscale adds an endpoint.
• Safe-by-default writes. Every tool declares readOnlyHint / destructiveHint / idempotentHint so clients can skip confirmation on reads and require it on mutations. A skill that shells out to the CLI can't express that.
• Real tests. 1900+ unit tests covering every tool's input validation, API shape, and error handling. Plus an opt-in live-tailnet integration suite (RUN_INTEGRATION_TESTS=1 + a tailnet API key) for shape-drift detection. Most skills are short markdown prompts without their own test layer — if the vendor changes output format, nothing catches it for you.

If you already have a skill that covers your 10% of Tailscale workflows, great — keep it. The MCP is for the other 90%.

What about Tailscale's own MCP endpoints and skill? As of 2026-09-19 they solve different problems, and nothing here duplicates them. Tailscale's official MCP tools are two alpha built-in connectors inside Aperture: Tailnet, whose Tailnet_provision_node returns a single-use auth key so an agent can join one new node after a person approves it, and Tailscale SSH, whose TailnetSSH_list_machines and TailnetSSH_run_command discover SSH-enabled machines and run one command on one of them. They require Aperture. tailscale/tailscale-skill is an alpha, knowledge-only skill — reference material that teaches an agent to curl the v2 API, with no server of its own. Neither exposes the admin API as typed tools, so the overlap with this server is close to nil. Nor can Aperture front this one today: this server speaks stdio, and Aperture proxies only URL-addressable Streamable-HTTP or SSE servers. Both of those products are alpha, so treat the date on this paragraph as its expiry.

Trust signals

Fair critique from Reddit: a new repo claiming "actively maintained" with no visible tests is worth exactly zero trust. Here's what's actually verifiable:

• 1900+ tests (node --test) covering every tool's input validation, API shape, and error handling. Run npm test to see them pass locally.
• Local release flow via release.sh: lint + test + bump + tag + push + npm publish + MCP Registry publish, all from the workstation. No CI workflow to babysit.
• Dependabot alerts surface on this repo and get fixed, not ignored.
• Every tool verified against the live API. If it's in the tool list, it calls a real endpoint that exists in the current v2 API. No placeholder 404 tools.

Issues and PRs are triaged. File one if something is off — github.com/YawLabs/tailscale-mcp/issues.

Quick start

1. Set your API key

Get an API key from Tailscale Admin Console > Settings > Keys and set it where your MCP client will see it. The .mcp.json env block in step 2 works identically on every platform and is the option to prefer; to export it from a shell profile instead (~/.bashrc, ~/.zshrc, ~/.config/fish/config.fish):

macOS / Linux / WSL (bash, zsh):

export TAILSCALE_API_KEY="tskey-api-..."

fish:

set -Ux TAILSCALE_API_KEY tskey-api-...

Windows (PowerShell 5.1 and 7) — [Environment]::SetEnvironmentVariable persists it for the user, where $env: alone lasts only for the session:

[Environment]::SetEnvironmentVariable('TAILSCALE_API_KEY', 'tskey-api-...', 'User')

2. Create .mcp.json in your project root

macOS / Linux / WSL:

{
  "mcpServers": {
    "tailscale": {
      "command": "npx",
      "args": ["-y", "@yawlabs/tailscale-mcp@latest"]
    }
  }
}

Windows:

{
  "mcpServers": {
    "tailscale": {
      "command": "cmd",
      "args": ["/c", "npx", "-y", "@yawlabs/tailscale-mcp@latest"]
    }
  }
}

Why the extra step on Windows? On Windows, npx is a .cmd file, and Node 20+ refuses to spawn .cmd files directly. Wrapping with cmd /c is the standard workaround.

3. Restart and approve

Restart Claude Code (or your MCP client) and approve the Tailscale MCP server when prompted.

That's it. Now ask your agent:

"List my Tailscale devices that haven't been seen in the last 7 days"

"Summarize every ACL change in the audit log from yesterday"

"Draft an ACL rule that lets tag:ci reach tag:registry on port 5000 only"

Too many tools? Subset them.

97 tools is a lot. If you've already got a dozen MCP servers and your client is feeling heavy, trim what this one exposes. Three knobs, combinable:

The env blocks below show only the variable under discussion. Your credentials come from the environment, as set in Quick start — keep them in your shell profile rather than in the client's JSON config, which is world-readable on most systems and easy to commit by accident.

Option 1: TAILSCALE_PROFILE (preset, easiest)

{
  "env": {
    "TAILSCALE_PROFILE": "core"
  }
}

• minimal (20 tools) — status, devices, audit. Observe the tailnet, read the audit log.
• core (52 tools) — adds acl, dns, keys, users. The day-to-day admin surface.
• full (97 tools, default) — everything. Same as omitting the env var.

Option 2: TAILSCALE_TOOLS (explicit group list)

{
  "env": {
    "TAILSCALE_TOOLS": "devices,acl,dns,audit"
  }
}

Comma-separated group names. Overrides TAILSCALE_PROFILE when both are set — use this when the presets aren't quite right.

Valid group names: status, devices, acl, dns, keys, users, tailnet, webhooks, posture, audit, invites, services, log-streaming. The local-cli group is also available, but only when TAILSCALE_LOCAL_CLI=1 is set — see Local CLI integration.

Option 3: TAILSCALE_READONLY (drop mutations)

{
  "env": {
    "TAILSCALE_PROFILE": "core",
    "TAILSCALE_READONLY": "1"
  }
}

Set to 1 or true to drop every tool without readOnlyHint: true. Stacks with TAILSCALE_PROFILE or TAILSCALE_TOOLS as an intersection — combine for maximum minimalism.

Confirming what loaded

The server logs the active filter to stderr on startup:

@yawlabs/tailscale-mcp v0.12.0 ready (20 tools, profile=minimal, readonly)

When both TAILSCALE_PROFILE and TAILSCALE_TOOLS are set, TAILSCALE_TOOLS wins. The banner marks the profile as overridden so the precedence is obvious at a glance — no need to guess which filter actually applied:

@yawlabs/tailscale-mcp v0.12.0 ready (22 tools, profile=core (overridden by TAILSCALE_TOOLS), groups=devices,acl)

The "(overridden)" marker only fires for substantive profiles (minimal / core); profile=full is a no-op preset, so it's shown without the marker when TAILSCALE_TOOLS is also set.

If you don't set any filter, startup prints a tip pointing you at the profiles.

And how the agent knows

Everything above is stderr -- your MCP client's log. The model never sees it, so a withheld tool and a tool that was never built look identical from the agent's side. That is how an agent ends up working around a restriction instead of reporting it.

tailscale_tool_groups closes that gap. It is always registered, whatever the filters say, and answers the question in-band:

> "Why can't you delete that device?"

  tailscale_tool_groups({ toolName: "tailscale_delete_device" })

  {
    "tool": "tailscale_delete_device",
    "available": false,
    "group": "devices",
    "kind": "write",
    "reason": "TAILSCALE_WRITE_GROUPS is set to \"dns\", which does not grant writes here",
    "toEnable": "add \"devices\" to TAILSCALE_WRITE_GROUPS (e.g. \"dns,devices\")"
  }

It separates the three cases an agent otherwise cannot tell apart:

Called with no arguments it lists every group with its availability and, for anything withheld, the exact environment change that would restore it. It reads no network and needs no credentials, so it works even when the server is misconfigured.

Scoping writes to areas

TAILSCALE_WRITE_GROUPS names the areas an agent may write to. Everything else stays readable:

{
  "env": {
    "TAILSCALE_WRITE_GROUPS": "devices,keys"
  }
}

That serves all 41 read tools plus the 18 writes in devices and keys, and withholds the other 38 writes — the ACL, DNS, users, tailnet, webhooks, posture, services, invites, org-tailnets and log-streaming writes are simply not registered. Unset means no write gate, which is the shipped default.

Read this first: this filters the tool list, not your API token. The server still holds one credential with full tailnet authority in every configuration. An agent that also has a shell can curl api.tailscale.com with that same token and do everything this knob withheld. Scope the Tailscale OAuth client itself to the areas you actually need (scopes per tool group) — that bound survives outside this process; this one does not. TAILSCALE_WRITE_GROUPS is the low-friction complement to credential scoping, not a replacement for it.

What a grant actually contains

Group names are the same ones TAILSCALE_TOOLS uses. Writes per group:

status, audit and the opt-in local-cli group contain no writes at all, so granting them does nothing.

devices is the widest grant, and the one most likely to be set. In a compose file write=devices reads like "device admin", but it hands over delete_device, set_devices_authorized, expire_device, deauthorize_device, set_device_routes, set_device_tags and update_device_key alongside rename_device. There is no finer setting: an honest "safe subset" of devices is rename_device alone, and a knob whose useful value is one tool is not a knob.

Three grants are tailnet-admin-equivalent

keys, users and acl are not blocked — CI key rotation legitimately needs keys — but grant them knowing:

• keys — tailscale_create_key mints an OAuth client with whatever scopes the caller asks for, including policy_file and all. That credential outlives the agent's session and is not subject to this or any other setting here.
• users — tailscale_update_user_role accepts owner.
• acl — tailscale_update_acl rewrites policy for every principal in the tailnet.

The server prints this on startup when your grant includes one of them.

What it does and does not bound

It bounds where an agent may write. It does not bound severity within a granted area: inside a granted group, writes run unattended, including the grant-direction ones. Pair it with TAILSCALE_REQUIRE_APPROVAL=1 when a human is at the keyboard — but understand that for an unattended agent that pairing contributes nothing, because a never-prompt client denies those calls rather than prompting.

Precedence, and what happens when you get it wrong

Names are case-sensitive, matching TAILSCALE_TOOLS.

The upgrade contract: upgrading this package can never widen the set of areas an agent may write to. A new group is in nobody's grant until a human types its name. It can, however, add tools inside an area you already granted — a pinned test makes that a reviewed line in the diff rather than a silent change.

The startup banner shows what applied:

@yawlabs/tailscale-mcp v0.19.0 ready (59 tools, write=devices,keys)

Requiring approval on irreversible tools

readOnlyHint / destructiveHint are advisory — the MCP spec says clients MUST treat annotations as untrusted, and most don't gate on them. TAILSCALE_REQUIRE_APPROVAL=1 adds a stronger signal that supported clients enforce:

{
  "env": {
    "TAILSCALE_REQUIRE_APPROVAL": "1"
  }
}

Nine tools are then advertised with _meta["anthropic/requiresUserInteraction"], which forces a confirmation prompt even when an allow-rule would otherwise auto-approve the call:

The line drawn is "this server cannot undo it with information you still hold", which is narrower than the 23 tools annotated destructiveHint: true. tailscale_suspend_user is deliberately excluded — tailscale_restore_user reverses it. So are tailscale_deauthorize_device (reversed by tailscale_authorize_device) and the replace-all setters, which all have a get_* counterpart you can read before writing.

Opt-in on purpose, and read this before turning it on. In a client mode that never prompts (an unattended agent, a CI run), the flag causes those calls to be denied rather than run. That is the right default when a human is at the keyboard and the wrong one when nobody is, so it stays off unless you set it.

Requires a client that honors the annotation; Claude Code added support in v2.1.199. Clients that don't recognize it ignore it, so setting the variable is never worse than leaving it off.

Separately and always on, the tools whose response size scales with the tailnet rather than with the request — tailscale_list_devices, tailscale_list_users, tailscale_get_acl, tailscale_diff_acl_access, tailscale_get_audit_log, tailscale_get_network_flow_logs, tailscale_local_status — declare _meta["anthropic/maxResultSizeChars"], so a large-but-legitimate result stays inline instead of being truncated into a file reference the agent has to read back mid-task. The last of those is the one entry behind an opt-in: it declares the cap whenever TAILSCALE_LOCAL_CLI=1 registers it, and is absent entirely otherwise.

Using with mcp.hosting / mcph

If you run this server through mcp.hosting (via the @yawlabs/mcph local agent), the two filtering layers compose cleanly:

1. Server-side — TAILSCALE_PROFILE / TAILSCALE_TOOLS / TAILSCALE_READONLY reduce the tool surface before mcph sees it. The unloaded tools aren't registered at all.
2. Client-side — mcph's mcp_connect_activate({ tools: [...] }) filters further for what appears in tools/list. Tools not in that list stay reachable via mcp_connect_dispatch, so you don't lose capability.

Recommended pattern for mcph users: set TAILSCALE_PROFILE=core (or narrower) in your mcp.hosting server config, then let mcph handle per-conversation activation on top. The server stays lean by default, and mcp_connect_dispatch covers the long-tail tools for ad-hoc needs.

Authentication

API key (simplest): Set TAILSCALE_API_KEY in your shell or MCP config.

OAuth (scoped access): For fine-grained permissions, set TAILSCALE_OAUTH_CLIENT_ID and TAILSCALE_OAUTH_CLIENT_SECRET instead. Create an OAuth client at Tailscale Admin Console > Settings > Trust credentials, with the scopes listed below for the tool groups you load.

The server checks for an API key first, then falls back to OAuth. If neither is set, tools return a clear error telling you what to configure — the server still starts, so your MCP client doesn't loop restarting.

Tailnet: Uses the credential's own tailnet (-) automatically, which is what most setups want. To name one explicitly, set TAILSCALE_TAILNET to the Tailnet ID shown under Settings > General in the admin console — it looks like T1234CNTRL. Tailnets created before October 2025 can still use their legacy organization name; newer ones have no such name to use. Per the OpenAPI spec, the Tailnet ID is the preferred identifier either way.

TAILSCALE_OAUTH_TAILNET — target an API-only tailnet (one created by tailscale_create_org_tailnet). Those tailnets are not reachable with a plain client-credentials exchange: you authenticate with an OAuth client belonging to the creating tailnet (all scope) and the target rides on the token request. Set this to the new tailnet's id. Deliberately separate from TAILSCALE_TAILNET so the default token exchange is unchanged for everyone else. If you set this, leave TAILSCALE_TAILNET unset (or -) so tool requests follow the token — pointing the two at different tailnets makes every tailnet-scoped tool return 403, and the server warns about it at startup.

OAuth scopes by tool group

The scopes an OAuth client needs for each TAILSCALE_TOOLS group. Grant the Read column for the groups you load, and add the Write column for the ones you let it change; Notes lists what a group needs beyond that. all:read (every read scope) and all (everything) are the broadest grants there are.

Scopes are taken from Tailscale's OpenAPI spec as of 2026-09-19. The DNS row, which the spec omits, and the device scopes in the acl row come from the trust credentials doc. All of it is read from the documentation, not observed against a tailnet.

Reliability and debugging

429 and gateway-error retry (built-in). HTTP 429, 502, 503 and 504 are retried up to 3 times on the idempotent methods (GET, PUT, DELETE), honoring the Retry-After header (both seconds-integer and HTTP-date forms). Falls back to exponential backoff with jitter, capped at 30s per wait. No env var needed — this is on by default. Workflows like "rotate every key older than 90 days" no longer fail mid-loop on Tailscale's per-tenant rate limits, and a gateway blip no longer fails a whole tool call: per the OpenAPI spec, 504 is documented on every device and service operation with the message "request took too long to process, please try again later", and 502 on the log reads. HTTP 500 is not retried — it means the server failed to process the request, not that something in front of it gave up. POST and PATCH are never retried, on any status. A DELETE that retried past a gateway error and then got a 404 says so in its error, because the attempt that timed out may already have deleted the resource.

TAILSCALE_DEBUG=1 — log every HTTP method, URL, status, and elapsed time to stderr. Authorization headers are never logged. Use this when a tool returns an unexpected error and you want to see the actual request that went out. Example:

[tailscale-mcp] GET https://api.tailscale.com/api/v2/tailnet/-/devices
[tailscale-mcp]   <- 200 (148ms)

TAILSCALE_MAX_CONCURRENT=N — cap in-flight API requests at N. Default is unlimited (no behavior change for users who don't opt in). Useful when an agent fans out aggressively against a tailnet that has stricter limits than the per-call retry can absorb.

TAILSCALE_REQUEST_BUDGET_MS=N — total wall-clock budget per request, including retries and their sleeps. Default 90000 (90s). When the next retry's predicted wall time would exceed the budget, the call surfaces the error immediately instead of holding the line. For a gateway 5xx that prediction also charges the duration of the attempt that just failed: a 504 arrives only after the gateway has already waited, so the retry most likely costs the same again, and spending the rest of the budget on it would leave your client with silence instead of the 504. A call that retries past a gateway 5xx is also held to half this budget from that point on — 45s by default, under the 60s low end of the usual MCP client timeout — since those attempts cost gateway wait time rather than backoff, and a chain of them can outlast the client while a 429 chain cannot. Raising this value raises that ceiling with it; a 429 chain keeps the whole budget either way. Tune lower if your MCP client has a tighter outer timeout. Non-idempotent methods (POST, PATCH) are never retried — those return immediately regardless of budget.

TAILSCALE_RETRY_BASE_DELAY_MS=N — base delay for the exponential backoff between retries; attempt N waits base * 2^N (capped at 30s, plus jitter). Default 1000 (1s), so a fully-exhausted retry chain spends roughly 1s + 2s + 4s sleeping. Pairs with TAILSCALE_REQUEST_BUDGET_MS: lowering the budget on its own doesn't get you more retries, it just makes the default backoff exhaust the budget sooner and give up. Shrink both if you want "retry hard, fail fast". A server-supplied Retry-After header always wins over this value.

TAILSCALE_EXTRA_WEBHOOK_EVENTS=eventA,eventB — opt-in escape hatch for webhook event types Tailscale ships after the latest release of this package. The webhook tools validate subscriptions against a strict static catalog so typos and stale event names fail fast with a clear error; if you need a brand-new event before the catalog catches up, list it here (comma-separated) and the schema will accept it. The two category subscriptions (categoryTailnetManagement, categoryDeviceMisconfigurations) are in the catalog, so they need no entry here. Please also open an issue so the static list catches up.

TAILSCALE_EXTRA_POSTURE_PROVIDERS=providerA,providerB — the same escape hatch for device-posture integration providers. tailscale_create_posture_integration validates provider against a static list (falcon, fleet, huntress, intune, jamfpro, kandji, kolide, sentinelone); if Tailscale adds one before this package catches up, list it here rather than waiting for a release. This field used to be a closed enum, which made a newly-supported provider uncreatable rather than merely unvalidated.

Friendlier error messages. JSON error bodies of the form {"message":"..."} or {"error":"..."} are unwrapped before display, so you see the prose explanation instead of raw JSON. When the body also carries a data array — which the ACL endpoints use to report a failing policy test — it is rendered under the message, so a rejected policy says which user and which assertion failed instead of just test(s) failed. 401s still get the full multi-line auth-error formatter (with the Windows env-var hint when applicable).

Local CLI integration (opt-in)

Most tools talk to the Tailscale v2 admin API — they describe the tailnet. Sometimes you want to ask about this machine's view: is it actually connected? What DERP region is it on? How far is my-laptop from here? Those answers come from the local tailscale binary, not the admin API.

Set TAILSCALE_LOCAL_CLI=1 (in your shell or .mcp.json env block) to add 6 read-only diagnostic tools:

Requirements: the tailscale binary has to be findable. It's looked up on PATH first, then at the default install paths below, and TAILSCALE_BINARY overrides both with an absolute path of your choosing.

The MCP server doesn't need root to run these — they're all diagnostic, not state-mutating. Operations that would need elevation (tailscale up, set --advertise-routes, lock sign) are deliberately not exposed.

When opt-in is on, the startup banner reflects it: @yawlabs/tailscale-mcp v0.13.3 ready (103 tools, local-cli=on) — the 6 local CLI tools are additive on top of the default 97.

Resources (4)

MCP Resources expose read-only data clients can browse without a tool call.

Tools (97 + 6 opt-in)

GitOps: deploy ACLs from CI

For the simple "deploy ACL from git on merge" workflow, you don't need an MCP server or an agent — use the built-in CLI:

# PR check: validate the proposed policy without touching the tailnet
npx -y @yawlabs/tailscale-mcp@latest validate-acl tailscale/acl.json

# On merge: ETag fetch + validate + deploy with If-Match, fail-closed at every step
npx -y @yawlabs/tailscale-mcp@latest deploy-acl tailscale/acl.json

Works in any CI system. Set TAILSCALE_API_KEY as an env var; TAILSCALE_TAILNET is optional — leave it unset to act on the key's own tailnet, or set it to the Tailnet ID to name one explicitly. Both commands exit non-zero on any failure; deploy-acl refuses to deploy without an ETag (so a concurrent Admin Console edit can never be silently clobbered) and reports a 412 as a concurrent-edit conflict you resolve by re-running.

When validation reports a failing policy test, the CI log names the user and the assertion (For user user1@example.com: / Errors found:), the same detail upstream's gitops-pusher prints. Validation warnings — a SCIM group that is not syncing, for instance — fail the run too, matching gitops-pusher and Tailscale's own Go client; their text is printed alongside so you can see what was flagged.

A complete GitHub Actions workflow — validate on PR, deploy on merge:

name: tailscale-acl
on:
  pull_request:
    paths: ["tailscale/acl.json"]
  push:
    branches: [main]
    paths: ["tailscale/acl.json"]

jobs:
  acl:
    runs-on: ubuntu-latest
    env:
      TAILSCALE_API_KEY: ${{ secrets.TAILSCALE_API_KEY }}
      TAILSCALE_TAILNET: T1234CNTRL # Tailnet ID from Settings > General; or omit: defaults to the key's tailnet
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: 22 }
      - name: Validate ACL
        if: github.event_name == 'pull_request'
        run: npx -y @yawlabs/tailscale-mcp@latest validate-acl tailscale/acl.json
      - name: Deploy ACL
        if: github.event_name == 'push'
        run: npx -y @yawlabs/tailscale-mcp@latest deploy-acl tailscale/acl.json

For reproducible deploys, replace @latest with a pinned version.

If you hand-roll this with curl instead: Tailscale's ACL endpoint only returns the ETag header on GET, not HEAD. A curl -I (HEAD) ETag fetch silently yields an empty value — and an empty If-Match either deploys unguarded (clobbering concurrent edits) or trips your guard and fails the deploy. Fetch the ETag with a GET (curl -fsS -D - -o /dev/null ...), and fail the job if it comes back empty. The CLI above does all of this for you.

Optional: Lock the Admin Console to prevent manual edits that drift from git. Ask your agent:

"Set aclsExternallyManagedOn to true and aclsExternalLink to our repo URL"

This shows a read-only banner in the Tailscale Admin Console pointing to your repo. Use the MCP for reads and investigations, and let CI handle the deploy.

Requirements

• Node.js 20.11+ to run the server (22+ to develop — the test script passes a glob to node --test, supported from Node 21)
• A Tailscale API key or OAuth client credentials

Running on oam.js (optional)

oam.js runs this server unmodified, and the tailscale-mcp command only ever uses the latest oam release, currently 0.15.2. Verified against oam 0.15.2: full MCP handshake, all 97 admin-API tools plus tailscale_tool_groups, all 4 resources, identical error responses, and a clean stdout protocol stream — from the shipped bundle and straight from the TypeScript source with no build step.

oam 0.15.2 is the minimum. A floor matters here: releases before 0.9.0 ran child_process.execFile arguments through a shell, re-splitting them on whitespace and executing shell metacharacters inside an argument, and this server shells out to the tailscale binary across its local-CLI tools, so that was a reachable bug rather than a theoretical one.

How the tailscale-mcp command (bin/tailscale-mcp.mjs) picks a runtime:

• TAILSCALE_MCP_RUNTIME=auto (the default) — if a client already launched it with oam run on oam 0.15.2 or newer, the server runs in that process. Otherwise it uses OAM_BIN when that is 0.15.2 or newer, else asks every oam binary it can find — %LOCALAPPDATA%\oam\bin then ~/.oam/bin on Windows, ~/.oam/bin elsewhere, then PATH — for its version and uses the newest at or above the floor (on a tie the installed copy wins). With none, it runs on Node. An oam host older than 0.15.2 never serves the server itself: it hands off to the newest usable oam, or to Node on PATH, or exits with an error when there is neither. Whenever it looks for an oam, stderr names an OAM_BIN that was passed over and why; the oam binaries it found and passed over are named, each with its reason, only when no usable oam turns up.
• TAILSCALE_MCP_RUNTIME=oam — the same, but exit with an error instead of falling back to Node.
• TAILSCALE_MCP_RUNTIME=node — always Node: in-process under npx, handed off to Node on PATH when a client launches the command with oam run.

The value is case-insensitive; anything else is warned about on stderr and treated as auto. On Windows only oam.exe counts: an oam.cmd / oam.bat shim is never run, and it is named on stderr only when no usable oam turns up.

Sandboxing (opt-in)

Set TAILSCALE_MCP_SANDBOX=1 to run under oam's --permission model: network restricted to api.tailscale.com -- the only host the bundle contacts, including the OAuth token exchange -- and filesystem denied. Child-process stays granted because the local-CLI tools shell out to the tailscale binary, which is also why PATH remains in the environment allow-list.

It is opt-in rather than default because a wrong grant does not fail loudly. oam denies a non-granted environment variable by making it absent from process.env rather than throwing, so an under-granted TAILSCALE_API_KEY reads as "unauthenticated" rather than "denied". The env allow-list in the launcher is derived from what the shipped bundle actually reads -- if you add a new process.env lookup, extend that list with it.

The sandbox is applied by the tailscale-mcp command, which spawns a fresh oam for it -- even when a client launched the command with oam run -- because --permission is a process-level flag. If it finds no usable oam to spawn, TAILSCALE_MCP_RUNTIME=auto still starts the server, without the sandbox; set TAILSCALE_MCP_RUNTIME=oam to make that an error. TAILSCALE_MCP_RUNTIME=node runs on Node, so it never applies the sandbox.

{
  "mcpServers": {
    "tailscale": {
      "command": "oam",
      "args": ["run", "/path/to/tailscale-mcp/dist/index.js"]
    }
  }
}

Measure startup on your own hardware. An MCP client cold-starts this server once per session, so startup is the cost that actually gets paid, and on the machine this was measured on node won it — 437ms vs 1554ms for oam run over 10 warmed runs (an earlier 5-run round showed 326ms vs 427ms; the box was busy, so treat the magnitude as noisy and the direction as the finding). Those runs used an oam that predates the 0.15.2 floor and have not been repeated since, so do not read them as a current ranking.

The published tailscale-mcp command prefers the newest usable oam it finds (see above). Without oam that costs almost nothing: discovery is file-existence checks only, never a subprocess, and the fallback runs the server inside the Node process npm already started. With oam installed, though, the command boots Node, runs --version on every oam binary it found to pick the newest, and only then boots oam, so it is always slower than pointing your client at a runtime directly — the config above for oam, node /path/to/tailscale-mcp/dist/index.js for Node. TAILSCALE_MCP_RUNTIME=node skips oam entirely.

Two places oam does win for this repo, both opt-in and neither touching the npm package:

• npm run check:oam — type-checks via oam check (tsgo, TypeScript 7 native). Measured 4015ms against 7680ms for tsc --noEmit, same clean result. npx tsc --noEmit remains the portable default.
• npm run build:binary:oam — builds the standalone binary via oam compile instead of Node SEA. Measured ~57.7 MB against ~73.6 MB for the Node SEA carrier before its blob is injected. Writes to the same bin/<platform>-<arch>/ path as npm run build:binary, so the release staging script consumes either unchanged — run one or the other. If you redistribute that binary it embeds oam's runtime, so ship oam's LICENSE, NOTICE and THIRD_PARTY_LICENSES.md with it.

The source stays runtime-agnostic on purpose: no oam: imports anywhere, and tests stay on node:test. That is what keeps the Node fallback real rather than nominal. Note that any oam invocation writes a bytecode cache to oam/ in the working directory — already in .gitignore.

Contributing

Contributions welcome. See CONTRIBUTING.md for the PR workflow and AI-agent guidelines. Please open an issue to discuss before a PR for anything beyond a typo fix.

git clone https://github.com/YawLabs/tailscale-mcp.git
cd tailscale-mcp
npm install
npm run lint       # Biome check
npm run lint:fix   # Auto-fix
npm run build      # tsc + esbuild bundle
npm test           # node --test (full suite)

For integration testing against your own tailnet: set TAILSCALE_API_KEY and run node dist/index.js.

Security

Found a vulnerability? See SECURITY.md — please use GitHub's private vulnerability reporting, not a public issue.

License

MIT