Runpod MCP server
Runpod MCP server is the official control-plane MCP server for managing Pods, Serverless endpoints, templates, network volumes, and related cloud resources over hosted HTTP or local stdio.
Project overview
Runpod MCP server is the official Runpod control-plane MCP server. It exposes Pods, Serverless endpoints, templates, network volumes, GPU capacity, and related account operations to MCP-capable agents through hosted HTTP with OAuth or local npx over stdio. It fits teams that want conversational access to Runpod operations, but it acts with the credential's full permissions, so every create, update, delete, or cost-bearing action should retain human approval.
Core capabilities
Pods and GPU capacity
Tools cover the Pod lifecycle, GPU types, and availability so an agent can inspect capacity before a user approves provisioning or changes.
Serverless endpoints and jobs
Manages endpoints, releases, and builds and submits sync, async, or streaming jobs; the docs warn that large outputs can exceed context limits.
Templates, volumes, and image access
Supports templates, network volumes, registries, and AWS ECR delegation. The docs prefer delegation for ECR because it avoids stored long-lived credentials.
Hosted and local modes
Hosted mode uses HTTP and OAuth, while local mode uses stdio and RUNPOD_API_KEY; both expose the same core control tools.
Task playbooks and hosted ALP
Both modes expose ten runpod://skills/ resources; hosted mode alone adds feedback, private journal, and question-collection tools.
Setup and connection
Start with the guided installer, choose the target client and configuration scope, then select hosted OAuth or a local API key. After setup, perform only a read-only identity check before enabling any cost-bearing or destructive operation.
Let an AI Agent install it
Send this prompt to Codex, Claude Code, or another AI agent that can work with your local environment.
Help me install Runpod MCP server from https://github.com/runpod/runpod-mcp. First read README.md, docs/configuration.md, docs/agent-learning-protocol.md, LICENSE, and the v4.0.0 Release, and note that the repository has no standalone SECURITY.md. Check that Node.js is at least 20, then ask whether I want hosted OAuth or a local API key and whether configuration should be user-scoped or project-scoped. Never display the API key in output, logs, or chat, and do not overwrite existing MCP configuration without confirmation. For hosted mode, remind me to select the correct Runpod account or team in the browser and explain anonymous analytics plus `X-Runpod-Analytics: off`. For local mode, confirm where the client will store RUNPOD_API_KEY. After configuration, run only an MCP connection check and a read-only Pod listing. Ask again before creating, updating, scaling, running a job, or deleting any resource.- Node.js 20 or later
- A Runpod account
- A Runpod API key for local mode
- An MCP client with HTTP/OAuth or stdio support
- Approval for client configuration writes and cloud-resource operations
npx @runpod/mcp-server@latest add- 1Check runtime and target client
Confirm Node.js 20+, record the MCP client and user or project scope, and back up existing configuration.
- 2Choose hosted or local mode
Choose hosted OAuth to avoid storing a key locally, or local stdio when you need local execution and no hosted analytics.
- 3Run guided or exact setup
Use `npx @runpod/mcp-server@latest add`, or apply the documented hosted HTTP or local stdio command for Claude Code.
- 4Verify connection and identity
Inspect the MCP server list, then perform a read-only Pod listing and verify the intended account or team; switch profiles and reauthorize if it is wrong.
- 5Add an approval boundary
Separate reads from writes. Show parameters and likely effects, then wait for approval before provisioning, updating, running cost-bearing jobs, or deleting resources.
First inspect the client's MCP server list; Claude Code can run `claude mcp list`. Then request only a Pod listing and verify that the returned resources belong to the intended account or team. Do not use create, update, job, or delete operations as the first test.
- The guided install command is `npx @runpod/mcp-server@latest add`; removal is `npx @runpod/mcp-server@latest remove`.
- The hosted Claude Code command is `claude mcp add --transport http runpod -s user https://mcp.getrunpod.io/`.
- For project-scoped Claude configuration, replace `-s user` with `-s project`.
- Hosted OAuth uses the Runpod profile selected in the browser at approval time; reauthorize after switching profiles.
Use cases
GPU capacity and Pod inventory
Summarize current Pods, GPU types, regions, and availability to support a human deployment choice.
Serverless releases and jobs
Inspect endpoints, releases, and builds, submit approved jobs, and route very large results to files or object storage.
Template and network-volume management
Inspect templates, network volumes, and image access settings and prepare a reviewable change plan.
Runpod operations playbooks
Read task resources under runpod://skills/ for official procedural context around deployment, troubleshooting, and resource management.
Assessment
The public repository's main strengths are official ownership, two deployment paths, a generated control-plane tool surface, and explicit documentation for identity selection, analytics, and large outputs. The tradeoffs are equally clear: credentials carry full permissions, tools can create billable resources, v4 removed legacy interfaces, and hosted mode includes data-collection features absent locally. This assessment is based on the README, configuration and ALP documents, Release, source layout, and test tree; it does not claim a live account test or long-term use.
Why it may be useful
- Officially maintained by Runpod with tools synchronized to the REST v2 specification
- Exact commands for hosted OAuth and local stdio deployments
- Published boundaries for profile selection, analytics, permissions, and large output
- Reviewable Apache-2.0 licensing, Releases, changelog, and test suite
What to know first
- The API key defines the full set of operations available to the agent
- v4 is incompatible with legacy v1 tools and several old names
- ALP is hosted-only and ask_question does not answer
- Installation and function were not tested in this cataloging run
README
Runpod MCP server
Overview
Runpod MCP server is the official control-plane MCP server for managing Pods, Serverless endpoints, templates, network volumes, and related cloud resources over hosted HTTP or local stdio. Official MCP server for interacting with the Runpod API.
Getting started
- Start with the guided installer, choose the target client and configuration scope, then select hosted OAuth or a local API key. After setup, perform only a read-only identity check before enabling any cost-bearing or destructive operation.
- The guided installer detects multiple MCP clients, lets the user choose hosted or local mode, and provides a matching remove flow.
- Hosted mode uses OAuth to obtain a session key for the Runpod account or team profile selected in the approving browser and does not store that key on disk.
- First inspect the client's MCP server list; Claude Code can run `claude mcp list`. Then request only a Pod listing and verify that the returned resources belong to the intended account or team. Do not use create, update, job, or delete operations as the first test.
Configuration
npx @runpod/mcp-server@latest add