runpod-mcp skill (Runpod): manage Runpod through the Runpod MCP server

Runpod's own skill that connects the Runpod MCP server and has a coding agent manage pods, endpoints, volumes and billing through typed tool calls.

Installs: 13.1K; on skills.sh
Stars: 47; on GitHub
Files: 3; besides SKILL.md
Updated: Oct 2026; last repo commit

runpod-mcp connects your coding agent to the hosted Runpod MCP server and makes it manage Runpod through that server's tools: pods, serverless endpoints and jobs, templates, network volumes, registry credentials, secrets, clusters, the GPU catalog and billing. It routes each request to a journey skill such as discovery, lifecycle-crud, pod-deploy or cost-audit.

It also sets the rules for every answer: facts come from tool reads, a deploy request counts as the go for that create, and only resources made in the conversation, or named by you, may be changed.

Good for

Developers who want an agent to deploy and operate Runpod pods and endpoints
Anyone auditing Runpod spend or checking GPU stock and prices
People diagnosing a pod that shows 0 GPUs, crashes or returns 502

Install runpod-mcp

Claude Code

Copy this prompt.

Install the runpod-mcp skill in Claude Code. Run this command:

/plugin marketplace add runpod/runpod-plugins-official
/plugin install runpod@runpod
/reload-plugins

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Pick one route, not both. Install the whole set, not only this skill: the runpod skills hand off to each other, and the worked golden-path examples live in the runpod router skill's folder. The plugin also registers the hosted Runpod MCP server, which stays inert until you sign in with /mcp.

From the project README. Source.

What's in the skill

When it loads

Manage Runpod infrastructure through the connected Runpod MCP server's tool calls (pods, serverless endpoints and jobs, templates, network volumes, registry credentials, secrets, clusters, the GPU/CPU catalog, and billing), and connect the server when it is missing (hosted or local npx). Load it before the first Runpod MCP tool call of a request: it routes the task to the journey skill that carries the procedure, owns the answer contract and the rules on which resources you may change, and defers to runpodctl for the terminal, file transfer, and SSH setup; template project and Hub listing authoring routes to runpod-build-template.

Sections in SKILL.md

Connect
The server is the source of truth
Use MCP vs runpodctl
Route by intent
Capability boundaries — state them, never fake them
Live reads before mutations
The answer contract
Source & docs

Key facts

Author: Runpod (vendor's own repo; the marketplace file and plugin manifest name Runpod as owner and author)
License: Apache 2.0 (LICENSE file, plugin manifest and SKILL.md agree)
Installs on skills.sh: 13,096 as of 2026-11-02
GitHub stars: 47 for the whole runpod-plugins-official repo, 2026-11-02
What is inside: 4 files including SKILL.md: 2 reference guides (connect.md, api-boundary.md) and 1 eval case. It bundles no scripts.
Last change to this skill: 2026-10-26
Part of: The runpod plugin (16 skills) in the runpod marketplace, version 1.8.0 in the manifest
Needs: A connected Runpod MCP server: OAuth sign-in with Runpod, or a RUNPOD_API_KEY

What it includes

Adds the hosted Runpod MCP server
Runs claude mcp add --transport http runpod -s user https://mcp.runpod.io/ and signs in with OAuth. With no user present it can send RUNPOD_API_KEY as a bearer header instead.
Runs code · Connect
Sends each request to the right journey skill
discovery for GPUs and costs, lifecycle-crud for create, inspect, update and delete, serverless-deploy, pod-deploy, pod-doctor, endpoint-ops and cost-audit.
Read · Route by intent
Rules every reply follows
Facts come from tool reads and are named in the answer. Diagnosis commits to one cause and one fix. A request to deploy or create is the go for that create, with the hourly price quoted.
Read · Answer contract
Only change what this conversation created or you named
A cleanup request does not extend that authority. For anything else the answer is an audit with the exact actions for you to take. Every write is followed by a read to confirm it.
Delete · Mutation limits
create-cluster and standing charges wait for your go
A cluster is never created until you have seen the shape and total hourly price and said go. A warm worker is created only after the standing charge is quoted per hour and per month.
Write · Explicit-go actions
Names what MCP cannot do
No SSH sessions, file transfer, local image builds or model downloads to volumes. SKILL.md sends those to runpodctl, SSH or the console.
Read · Capability boundaries
Calls the hosted server's initialize method
A curl command that posts an initialize request to mcp.runpod.io with your API key as a bearer header.
Runs code · Server version probe

From the official docs

Permissions

High risk

The agent gets tools that create, change and delete billable Runpod resources. Sign in with OAuth, read each create and delete before approving, and note that SKILL.md treats a deploy request as the go for that create.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

A deploy request counts as the go for that create
SKILL.md says a request to deploy, set up or create is the go, so the agent quotes the hourly price and makes the call without asking again. Only clusters and standing charges such as warm workers wait for an explicit go. Be specific about what you want created. github.com
It can delete resources it made or you named
SKILL.md lets the agent stop, update or delete resources created in this conversation, and the one resource you name. A cleanup instruction does not widen that. Name resources by id. github.com
The MCP server can use your API key as a bearer header
The key route puts your RUNPOD_API_KEY into the MCP client config. Prefer OAuth sign-in. github.com
The hosted server records anonymous tool use
The hosted server logs the tool name, status, duration and transport for each call, and drops tool arguments except on the guide and concept tools. Send the X-Runpod-Analytics: off header to opt out. The README says a locally run server sends none of this, though its docs search still goes to the hosted server. github.com
The plugin route installs all 16 skills and the Runpod MCP config
The marketplace file lists 16 skills for the runpod plugin. It also bundles .mcp.json, which registers the hosted Runpod MCP server at mcp.runpod.io, and a /runpod:migrate command. The git tree shows no hooks file in the plugin folder. github.com

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Talks to	mcp.runpod.io (hosted MCP server). api.runpod.io (v2 OpenAPI document)
Sign-in and keys	OAuth sign-in with Runpod, which SKILL.md says keeps no key on disk, or your RUNPOD_API_KEY sent as a bearer header when OAuth is not possible.
Scopes	Adds the hosted Runpod MCP server (mcp.runpod.io) to your client Creates, updates and deletes pods, endpoints, templates, volumes, registry credentials, secrets and clusters through MCP tools allowed-tools is Bash(claude mcp:*), so the agent can run claude mcp commands

Questions

What does the runpod-mcp skill do?

It connects your coding agent to the hosted Runpod MCP server and has it manage pods, serverless endpoints, templates, volumes, secrets and billing through typed tool calls.

How do I install runpod-mcp?

Use the install box at the top of this page. The plugin route also registers the hosted MCP server, which stays inert until you sign in. In Claude Code run /mcp, pick runpod, then Sign in with Runpod.

What is the Runpod MCP server URL?

https://mcp.runpod.io/ is the hosted server. SKILL.md also describes a local stdio server run from the npm package @runpod/mcp-server.

Does it need an API key?

Not for the OAuth sign-in. SKILL.md says OAuth signs in the MCP only, and that runpodctl or flash need a RUNPOD_API_KEY later. With no user in the loop it sends the key as a bearer header.

Can runpod-mcp delete my resources?

Yes, within limits. SKILL.md lets it stop, update or delete resources created in the conversation and the one you name. Cleanup instructions do not widen that, so unrelated resources get an audit and a list of actions for you.

Will it spend money without asking?

It can. SKILL.md treats a request to deploy or create as the go for that create and quotes the hourly price in the same reply. A cluster, and a warm worker standing charge, wait for your explicit go.

Is runpod-mcp safe?

It acts on real, billable infrastructure. Use OAuth, say exactly what to create, and read delete requests before approving. Our advice: start with read-only requests such as listing endpoints.

runpod-mcp or runpodctl?

Both drive the same Runpod API. Use runpod-mcp when the MCP tools are connected, and runpodctl for the terminal, SSH, file transfer and doctor setup.

Is this the official Runpod skill?

Yes. It is in Runpod's own runpod-plugins-official repo, and the plugin manifest names Runpod as author.

Information

updated Nov 2, 2026

Repository: runpod/runpod-plugins-official
Folder: plugins/runpod/skills/runpod-mcp

runpod/runpod-plugins-official skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 7

README.md · github.com
marketplace.json · github.com
SKILL.md · github.com
.mcp.json · github.com
README.md · github.com
connect.md · github.com
README.md · github.com