Runbook — Deploy / redeploy gcp-mcp-standalone

Prerequisites

• gcloud auth login and gcloud auth application-default login (Terraform needs ADC).
• Project bob-mcp-project-2026, billing on, Compute + Secret Manager APIs enabled.
• Service account mcp-broker-sa with secretAccessor on tailscale-auth-key and github-oauth-client-secret (created out-of-band).
• Secret Manager: a valid one-time, pre-authorized, non-ephemeral Tailscale auth key in tailscale-auth-key (see gcp-mcp-standalone/runbooks/rotate-tailscale-key).
• A Cloudflare Tunnel for bobsmcp.uk → localhost:8002; its token saved on the Mac: printf '%s' 'TOKEN' > ~/.cf-tunnel-token && chmod 600 ~/.cf-tunnel-token (never paste the token into a chat or shell history you don't control).
• Your Mac on the same Tailscale tailnet (Ansible rides it).
• terraform.tfvars holds only gcp_project_id — no secrets.

Steps

1. terraform plan -out=tfplan → review → terraform apply tfplan. Creates instance + deny-all firewall; first boot fetches the Tailscale key from Secret Manager and joins the mesh. Note: any change to metadata_startup_script forces VM replacement.
2. Confirm join: node gcp-mcp-broker active in the Tailscale admin console (or tailscale status on the Mac). If it registered as gcp-mcp-broker-1, an old machine entry existed — delete it and rename the new node.
3. Deploy: ansible-playbook -i ansible/inventory.ini ansible/install.yml -e cloudflare_tunnel_token="$(cat ~/.cf-tunnel-token)" Secret-bearing tasks are no_log; do not run with -v on a shared terminal anyway.

Verification

• On box: docker compose ps (broker Up), systemctl is-active cloudflared (active), sudo ss -tlnp | grep 8002 shows 127.0.0.1:8002 only.
• Public: curl -I https://bobsmcp.uk → 401 (alive, auth required); /.well-known/oauth-protected-resource/proxy/notion_api → 200; port probe against the public IP (nc -z <ip> 8002 / 22) fails.
• Client connects at https://bobsmcp.uk/proxy/<connector> (NOT the bare root).

Rollback

terraform destroy removes instance + firewall. Delete the machine from the Tailscale admin console (nodes are non-ephemeral). The Cloudflare tunnel survives for the next deploy; delete it in Zero Trust only if retiring the environment. VM replacement loses data/ (inbound + connector OAuth token DBs) and the generated .env — claude.ai must re-register and connectors must be re-authorized afterwards.