claudetools/wiki/patterns/tailscale-client-management.md

Pattern: Tailscale for Client Networks (MSP-managed)

Type: pattern Applies to: any ACG client that needs Tailscale (small offices especially) Companion script: tailscale-client-enroll.ps1 Last updated: 2026-06-06

---

Decision (TL;DR)

One tailnet PER client. Never merge a client into your own tailnet, and never share one tailnet across multiple clients. You hold an Admin/Owner seat on each client's tailnet, enroll their devices as tagged nodes via pre-auth keys pushed from GuruRMM, and bill on the client's own plan.

This gives clean isolation, clean offboarding, and zero reliance on a non-technical user doing anything.

---

Why per-client, not a shared tailnet

• Isolation / liability. Your tailnet carries ACG infra (the 100.x fleet, the vault host, coord API, GuruRMM server). One ACL slip and a client box - or malware on it - can reach your stuff or another client. Hard no.
• Billing. Client devices/users on your tailnet inflate and mix your bill. You want their cost on their ledger (or cleanly bundled into their managed-services fee).
• Offboarding. Ending an engagement = delete the tailnet / revoke your admin: one click. Untangling one client's nodes, ACLs, and keys out of a shared tailnet is error-prone and risky.
• IdP blast radius. Their Microsoft/Google identity changes shouldn't touch a tailnet that anyone else lives on.

---

Ownership and roles

1. Tie the tailnet to an identity that belongs to the client - their Microsoft 365 or Google Workspace (most ACG clients have one) as the login identity provider, or a dedicated admin account you hold on their behalf.
2. Add yourself as Owner/Admin. Tailscale roles: Owner, Admin, IT admin, Network admin, Auditor, Member. Admin is enough to manage devices/keys/ACLs.
3. Billing on the client's subscription. Two users fits the free tier functionally, but for a commercial client the Starter plan (~$6/user/mo) is the correct, supported footing - call it ~$12/mo, trivially bundled into their MSP fee.

---

Device enrollment (built for tech-inept users)

The only step in Tailscale that is hard for a non-technical user is the initial IdP login. Eliminate it entirely with tagged pre-auth keys - a tagged node is owned by a tag, not a user, so there is nothing to sign into and nothing that breaks when a password changes.

1. In the client's admin console: Settings -> Keys -> Generate auth key
   • Reusable (enroll both machines / tolerate RMM re-runs)
   • Pre-approved (skip manual device approval)
   • Tag: tag:<client> (e.g. tag:roberts)
   • Expiry ~90 days for the onboarding window; rotate after.
2. In GuruRMM, store the key as a secret/masked parameter and run tailscale-client-enroll.ps1 on each machine. It installs the MSI silently with TS_UNATTENDEDMODE=always (stays connected with no user logged in) and authenticates with the key. Zero user interaction; survives reboots.

You already have the GuruRMM agent on their boxes, so this is the whole job - no site visit, no walking a confused user through a login.

---

"Just see each other" ACL

Tailscale's default ACL is allow-all; the moment you set an ACL it becomes default-deny. This tag-scoped ACL lets the client's tagged machines reach each other and nothing else:

{
  "tagOwners": {
    "tag:roberts": ["autogroup:admin"]
  },
  "acls": [
    { "action": "accept", "src": ["tag:roberts"], "dst": ["tag:roberts:*"] }
  ]
}

With MagicDNS on, the two boxes reach each other by name (e.g. front-desk, back-office) with no IP juggling.

---

Subnet routing on Windows - the real story

• For two peers you do not need it. Each machine is a node with its own 100.x address and talks to the other directly. Skip subnet routing entirely.
• If you ever need to expose a whole LAN behind one box: modern Tailscale runs Windows subnet routers in userspace / netstack mode - no kernel IP-forwarding, no IPEnableRouter registry hack, and Windows automatically picks up the advertised routes. So it is not "hard" the way it used to be. The only real caveat is throughput: userspace routing is slower than Linux kernel-mode, so for a heavily used router drop in a cheap Linux box/appliance instead. Summary: easier than its reputation, just not the fastest.

---

Offboarding

Delete the client's tailnet, or revoke your admin seat and rotate the auth keys. Because the client is isolated in their own tailnet, the blast radius is exactly one client.

---

Multi-client management reality

There is no native single pane of glass across tailnets - each tailnet is its own admin console, so per-client tailnets means a separate login per client. For a handful of small clients that is fine. Record each client's tailnet name + admin identity + key rotation date in their wiki/clients/<client>.md. (Re-check Tailscale's MSP/partner program periodically for centralized multi-tenant tooling.)

---

Runbook - Robert's office (2 machines, "see each other")

1. Stand up the tailnet on Robert's identity; add yourself as Admin.
2. Turn on MagicDNS; set the ACL above (tag:roberts).
3. Generate a reusable + pre-approved auth key tagged tag:roberts.
4. In GuruRMM, add the key as a secret var and run tailscale-client-enroll.ps1 on both machines.
5. Confirm both show a 100.x IP in tailscale status; test reachability by MagicDNS name.
6. Record the tailnet name, your admin identity, and the key rotation date in Robert's client wiki article.

---

Sources

• Subnet routers - Tailscale Docs
• Kernel vs. netstack subnet routing & exit nodes - Tailscale Docs
• Install Tailscale on Windows with MSI - Tailscale Docs
• Keep Tailscale running when I'm not logged in - Tailscale Docs
• Auth keys - Tailscale Docs