AWS's managed MCP server: call any AWS API, run sandboxed Python against your account, and load AWS agent skills on demand — under IAM, CloudWatch and CloudTrail.
Runpod MCP Server
Summary
Runpod's official MCP server for driving GPU infrastructure — create and manage Pods, Serverless endpoints, templates and network volumes from an AI client.
Features
- Create, list, inspect and terminate GPU Pods by image and GPU type
- Manage Serverless endpoints, templates and network volumes
- Hosted endpoint with Sign in with Runpod OAuth — no API key stored on disk
- Guided installer detects your MCP clients and writes their config
- Local stdio mode with RUNPOD_API_KEY for bring-your-own-key setups
- Per-request tokens forwarded to the Runpod API and never persisted
Installation
Set up this MCP server in your favorite AI agent — copy a ready-made configuration below.
Description
The official Runpod MCP server puts the Runpod control plane behind an AI client, so provisioning a GPU box or scaling a Serverless endpoint becomes a sentence rather than a trip to the console. It manages Pods, Serverless endpoints, templates, network volumes and the rest of the account's resources.
Two ways to connect
- Hosted (recommended) — point the client at
https://mcp.getrunpod.io/and complete "Sign in with Runpod". The OAuth flow issues a session-scoped API key; nothing is written to disk and the server holds no credential of its own. In Claude Code that is one command:claude mcp add --transport http runpod -s user https://mcp.getrunpod.io/. - Local stdio —
npx -y @runpod/mcp-server@latestwithRUNPOD_API_KEYin the environment, for setups that would rather carry their own key.
Either way, npx @runpod/mcp-server@latest add runs a guided installer that detects which MCP clients you have, asks which to configure and writes the config itself; remove undoes it.
What it operates
Pods (create by image and GPU type, list, inspect, stop and terminate), Serverless endpoints (on the default v2 API these are image-based — supply an image and a GPU pool from list-gpu-types rather than a template), templates, and network volumes. Documentation covers REST v1 versus v2 selection and the templateId migration, private registry credentials versus ECR delegation, and how large tool outputs are handled.
Beyond the server
Runpod also publishes an official plugin marketplace that bundles this server with a router skill and five more — runpod-mcp, runpodctl, flash, runpod-usage and companion-clis — installable with npx skills add runpod/runpod-plugins-official across Claude Code, Codex, Cursor, Copilot, Windsurf, Cline, Gemini, opencode and others. In Claude Code the native plugin route wires up the hosted server with OAuth in one step.
Caution
The server acts with the full permissions of whatever key authenticates the request, and each request's token is forwarded to the Runpod API and never persisted server-side. Terminating a Pod is a real, billable, irreversible action — be deliberate with the destructive tools. Requires Node.js 18+. Apache-2.0.
Related MCP Servers
Google's official MCP server for driving the gcloud CLI from natural language, with allowlist and denylist controls over which commands an agent may run.
Google's open-source MCP server for enterprise databases — one binary that fronts AlloyDB, BigQuery, Spanner, Postgres, MySQL, Oracle, MongoDB, Snowflake and more.
Datadog's official managed MCP server — query logs, metrics, traces, incidents and dashboards from an AI agent, inside your existing Datadog permissions.
