Runpod MCP Server
Official Runpod MCP server: an agent manages Pods, Serverless endpoints, templates and network volumes from chat.
- 86 ★ on the repository
- Last commit
Broad access
Why this level
- The server acts with the key's full permissions: it creates billable Pods and deletes resources.
- Dangerous tools carry `destructiveHint`; confirmation depends on the client.
Install
- Works in
Manual install
npx @runpod/mcp-server@latest addThe installer detects clients and writes the configuration. Modes: hosted with Runpod sign-in or local with `RUNPOD_API_KEY`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The server acts with the key's full permissions: it creates billable Pods and deletes resources.
- Dangerous tools carry `destructiveHint`; confirmation depends on the client.
This is third-party code. Review the repository files before installing.
What it does
Official MCP server for Runpod, a GPU cloud. Claude Code, Cursor, Windsurf, VS Code and other clients use it to manage Pods, Serverless endpoints, templates, network volumes and container registries. Start, stop and restart of a Pod go through one `pod-action` tool, and `list-gpu-types` returns available GPU types. The server is hosted at `https://mcp.runpod.io/`: a client connects by URL and signs in with a Runpod account, with no key stored on disk. Alternatively the server runs locally from the `@runpod/mcp-server` npm package with your own `RUNPOD_API_KEY`. Every tool carries `readOnlyHint` and `destructiveHint` annotations that clients use to confirm dangerous calls.
Who it is for. Engineers who train and serve models on Runpod GPUs and want to spin up Pods and endpoints with chat commands.
Good fit when
- A Pod with an RTX 4090 is needed for an experiment and opening the console is a chore.
- A Serverless endpoint with a worker cap is created from a chat description.
- An agent reviews which Pods were left running and burning money.
- An image from a private registry is attached to a Pod.
- A local server in CI or a script runs with its own API key.
Not a fit when
- The agent cannot be trusted with a key holding full permissions on a Runpod account.
- GPU spend must not happen without manual confirmation of each step.
- The work runs on a different GPU cloud.
Example request
List all my Runpod Pods and create a Pod named test-pod on an NVIDIA GeForce RTX 4090 with one GPU.MCP
- Transport
- stdio, http
- Authentication
- OAuth
Environment variables
`RUNPOD_API_KEY`secret- Runpod API key for local runs. Not needed in hosted mode with Runpod sign-in.
Limitations
The server acts with the full permissions of the supplied key: it creates billable Pods and stops or deletes resources. Confirm destructive calls by hand, for example `delete-pod` and `purge-endpoint-queue`. Local use needs Node.js 20 or newer and an API key. In the 4.x line eight tools were renamed and `start-pod`, `stop-pod` and `restart-pod` merged into `pod-action`. If Runpod sign-in returns `403`, the firewall has temporarily blocked your IP. Wait up to an hour or send the key in an `Authorization` header.
How to disable. Run `npx @runpod/mcp-server@latest remove` or delete the `runpod` entry from the client configuration. Revoke the API key in the Runpod console.
FAQ
How do hosted and local modes differ?
In hosted mode the client connects to `https://mcp.runpod.io/` and signs in with Runpod over OAuth, with no stored key. In local mode the server starts through `npx` and `RUNPOD_API_KEY` sits in the client configuration.
Which clients does the installer configure?
`claude-code`, `cursor`, `windsurf`, `claude-desktop`, `vscode`, `opencode` and `pi`. `npx @runpod/mcp-server@latest add --list` shows detected clients and config paths.
How do I run it without the interactive prompt?
Name the clients with a flag, for example `npx @runpod/mcp-server@latest add --client cursor --local`. Without `--client` and without a terminal the installer exits with code 2 and configures nothing.
What if sign-in returns 403?
The firewall has temporarily blocked your IP after a burst of unauthenticated requests. The block lifts within an hour. To avoid a repeat, send your API key in an `Authorization: Bearer` header.
Does the server collect data?
The hosted server records one anonymous event per tool call. Opt out with the `X-Runpod-Analytics: off` header. Running locally sends nothing.
Discussion
No questions or reviews yet.
Related
MCP serversRunpod Plugins: GPU cloud for agents
Runpod Official Plugins
The Runpod plugin with 16 skills and an MCP server: an agent starts pods, deploys endpoints and audits GPU spend
Google's MCP server for deploying apps to Cloud Run from an agent, with service listing and log reading
MCPM
mcpm
A package manager and registry for MCP on all platforms and clients: finds, adds and configures MCP servers, remembers the source and updates
mcpd starts MCP servers from one .mcpd.toml file and serves them to agents and scripts as a plain HTTP API on a local port
A Node.js manager for MCP servers: it starts local servers, connects remote ones and gives a client one address, with a REST API for control
Official Render MCP server: services, deploys, logs, metrics and Postgres databases on Render
README in short
Tools are generated from the Runpod OpenAPI spec, so the set tracks the cloud API. The 3.x upgrade path is documented separately.