Google Cloud Run MCP — Anwendungsfälle, Installieren & Live-Demo

Warum nutzen

Hauptfunktionen

First-party — maintained by GoogleCloudPlatform
Deploy from local folder OR raw file contents (no Docker build needed)
Pull service logs for debugging without leaving the agent
Auth via ADC — the same gcloud auth you already use

Live-Demo

In der Praxis

cloud-run.replay ▶ bereit

0/0

Installieren

Wählen Sie Ihren Client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "cloud-run": {
      "command": "npx",
      "args": [
        "-y",
        "@google-cloud/cloud-run-mcp"
      ]
    }
  }
}

Öffne Claude Desktop → Settings → Developer → Edit Config. Nach dem Speichern neu starten.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "cloud-run": {
      "command": "npx",
      "args": [
        "-y",
        "@google-cloud/cloud-run-mcp"
      ]
    }
  }
}

Cursor nutzt das gleiche mcpServers-Schema wie Claude Desktop. Projektkonfiguration schlägt die globale.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "cloud-run": {
      "command": "npx",
      "args": [
        "-y",
        "@google-cloud/cloud-run-mcp"
      ]
    }
  }
}

Klicken Sie auf das MCP-Servers-Symbol in der Cline-Seitenleiste, dann "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "cloud-run": {
      "command": "npx",
      "args": [
        "-y",
        "@google-cloud/cloud-run-mcp"
      ]
    }
  }
}

Gleiche Struktur wie Claude Desktop. Windsurf neu starten zum Übernehmen.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "cloud-run",
      "command": "npx",
      "args": [
        "-y",
        "@google-cloud/cloud-run-mcp"
      ]
    }
  ]
}

Continue nutzt ein Array von Serverobjekten statt einer Map.

~/.config/zed/settings.json

{
  "context_servers": {
    "cloud-run": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "@google-cloud/cloud-run-mcp"
        ]
      }
    }
  }
}

In context_servers hinzufügen. Zed lädt beim Speichern neu.

claude mcp add cloud-run -- npx -y @google-cloud/cloud-run-mcp

Einzeiler. Prüfen mit claude mcp list. Entfernen mit claude mcp remove.

Anwendungsfälle

Praxisnahe Nutzung: Google Cloud Run

Deploy a small service to Cloud Run from the agent

👤 Backend engineers, solo founders shipping prototypes ⏱ ~10 min intermediate

Wann einsetzen: You have a Node/Python service ready. You want it live on Cloud Run without context-switching to the console.

Voraussetzungen

gcloud SDK installed + authenticated — gcloud auth login && gcloud auth application-default login
A GCP project with billing enabled — gcloud projects list

Ablauf

List projects, pick the right one

List my GCP projects. I want to deploy to the one tagged with 'staging'.✓ Kopiert

→ Project identified
Deploy the local folder

Deploy ./my-api as a Cloud Run service named 'api-staging' in us-central1. Allow unauthenticated invocations. Use the auto-detected buildpack.✓ Kopiert

→ Deploy completes, URL returned
Verify

Get the service details. Call the URL with a curl-style test. Confirm 200.✓ Kopiert

→ Working service URL

Ergebnis: A running Cloud Run service in 3-5 minutes, with the agent recording the URL for you.

Fallstricke

Unauthenticated flag exposes the service publicly — great for staging, bad for prod — For anything prod-adjacent, omit allow_unauthenticated and use IAM; front with Cloud Armor if user-facing
Each deploy creates a new revision, old ones linger and pin resources — Periodically clean up old revisions with gcloud, or configure revision retention

Kombinieren mit: fastmcp · fastapi-mcp

Diagnose a Cloud Run service from its logs

👤 On-call engineers ⏱ ~10 min intermediate

Wann einsetzen: A Cloud Run service is returning 500s. You want to see logs without opening the console.

Ablauf

Pull recent logs

Get the last 200 log entries for service 'api-prod' in us-central1. Filter to ERROR severity.✓ Kopiert

→ Error log entries
Find patterns

Group the errors by message template. Which one is most common? Include one sample stacktrace.✓ Kopiert

→ Top error with stacktrace
Propose remediation

Based on the stacktrace, point at the likely code path. Recommend a rollback to the previous revision if the errors are concentrated in the last deploy.✓ Kopiert

→ Root cause hypothesis + action

Ergebnis: From alert to rollback decision in 5 minutes.

Fallstricke

Log retention is short by default (30 days) — For anything you might need to look at later, export logs to BigQuery or GCS via Log Router

Kombinieren mit: sentry

Deploy an MCP server (FastMCP / fastapi-mcp) to Cloud Run

👤 Platform engineers hosting shared MCPs ⏱ ~30 min advanced

Wann einsetzen: You built an MCP server and want it hosted for team-wide use.

Voraussetzungen

Dockerfile or buildpack-compatible project — Cloud Run can use buildpacks for most languages

Ablauf

Deploy the server

Deploy ./my-mcp as Cloud Run service 'acme-mcp', region us-central1. Set env vars from my .env.prod. Require authentication.✓ Kopiert

→ Service deployed behind IAM
Grant IAM access

Grant role 'roles/run.invoker' on this service to [email protected] group.✓ Kopiert

→ Team can now invoke
Wire clients

Write the Claude Desktop config that connects via gcloud run services proxy or mcp-remote with IAM token.✓ Kopiert

→ Team has copy-paste setup

Ergebnis: A team-hosted MCP with GCP-native auth, same deploy model as your other services.

Fallstricke

SSE connections drop on Cloud Run's default 5-minute request timeout — Set --timeout=3600 on the service; MCP connections need longer idle tolerance

Kombinieren mit: fastmcp · fastapi-mcp · mcp-context-forge

Kombinationen

Mit anderen MCPs für 10-fache Wirkung

cloud-run + fastmcp

Write a Python MCP, deploy to Cloud Run in one session

Scaffold a FastMCP server exposing our pricing API, then deploy it to Cloud Run service 'pricing-mcp' with IAM auth.✓ Kopiert

cloud-run + sentry

Cloud Run errors → Sentry for incident triage

If our Cloud Run service's error rate spikes, pull Sentry's matching issue and correlate stack traces with the last deploy.✓ Kopiert

cloud-run + mcp-context-forge

Host ContextForge itself on Cloud Run to federate MCPs

Deploy mcp-contextforge-gateway to Cloud Run with Redis for federation state. Register our 3 existing MCPs as backends.✓ Kopiert

Werkzeuge

Was dieses MCP bereitstellt

Werkzeug	Eingaben	Wann aufrufen	Kosten
list-projects		Discover which project to target (local only)	free
create-project	name, project_id?	Bootstrap a new GCP project (local only)	free
list-services	project, region?	See what's running in a project/region	free
get-service	project, region, service	Inspect a specific service's config	free
get-service-log	project, region, service, filter?, limit?	Debugging — pull recent logs	free
deploy-local-folder	folder_path, service, region, project, allow_unauthenticated?	Deploy code you have locally (local only)	free (build) + per-request Cloud Run pricing
deploy-file-contents	files: {path,content}[], service, region, project	Deploy agent-generated code without local files	free (build) + per-request Cloud Run pricing

Kosten & Limits

Was der Betrieb kostet

API-Kontingent: GCP quotas — generous for Cloud Run deploy API
Tokens pro Aufruf: Log responses can be large — paginate or filter aggressively
Kosten in €: Cloud Run: free tier ~2M req/month, then ~$0.40 per million req + $0.000024 per vCPU-sec
Tipp: Set min-instances=0 for non-critical services — cold starts hurt but cost drops to near-zero at idle

Sicherheit

Rechte, Secrets, Reichweite

Minimale Scopes: roles/run.admin roles/iam.serviceAccountUser

Credential-Speicherung: Application Default Credentials via gcloud auth application-default login — no key files

Datenabfluss: Calls to *.googleapis.com

Niemals gewähren: roles/owner

Fehlerbehebung

Häufige Fehler und Lösungen

Permission denied on deploy

Your account needs roles/run.admin AND roles/iam.serviceAccountUser on the runtime service account. Check via gcloud projects get-iam-policy.

Build failed with 'buildpack not detected'

Add a Dockerfile, or a file that buildpacks recognize (package.json for Node, requirements.txt for Python, etc.).

Service deployed but returns 403

Either allow_unauthenticated wasn't set, or your invoking identity lacks roles/run.invoker. Decide: public or IAM-gated.

SSE MCP connections hang / drop

Increase --timeout=3600 and ensure min-instances=1 to avoid cold start disconnects.

Alternativen

Google Cloud Run vs. andere

Alternative	Wann stattdessen	Kompromiss
Vercel (for Next.js)	You're specifically hosting a Next.js app	Simpler DX for Next.js; narrower than Cloud Run
AWS App Runner / Lambda MCP	You're on AWS	Comparable serverless model, different ecosystem
Fly.io	You want global edge by default and simpler pricing	Smaller ecosystem; no official MCP yet

Mehr

Ressourcen

📖 Offizielle README auf GitHub lesen

🐙 Offene Issues ansehen

🔍 Alle 400+ MCP-Server und Skills durchsuchen