How Runpod’s agent skills and MCP server work together
An API tells an agent what it can do, not whether it should. See how skills, a concept graph and live-account tests keep agents right on Runpod.
All
You can hand Runpod setup, deploys and cleanup to your coding agent with one prompt. The prompt installs two things: skills, which are tested procedures for jobs like deploying a Pod, and a hosted MCP server, which gives the agent Runpod tools and current docs. Without them, an agent can work from memory and guess at endpoints, flags and defaults it learned from documentation that may be out of date. Every MCP tool carries read-only and destructive labels, so you decide which actions your agent can take without asking.
Run one prompt
Read agents.runpod.io and set up Runpod for me.The agent installs the skills, registers the MCP server at https://mcp.runpod.io/ and starts sign-in. It may ask whether to install for your user account or only the current project. Then approve sign-in in your browser, or use an API key if your client doesn’t support browser sign-in (the API key steps are in Run it from CI below).
Here is the flow in Claude Code, sped up. This run also asks the agent to deploy a Pod.
Claude Code reads agents.runpod.io, installs the Runpod plugin, signs in and deploys a Pod. Fast-forwarded from a few minutes.
If your agent isn’t on the supported list, it can follow the same instructions through the universal skills option, with a few more questions. Agents that load Claude-format plugins can instead install the official plugin from runpod/runpod-plugins-official, which carries the skills and the hosted MCP server. Slash commands such as /runpod:migrate may not carry over, but the runpod-migrate skill behind them does.
What skills add to an MCP server
Hyperscalers and neoclouds now ship MCP servers of their own, so the table compares what each approach gives your agent once it’s connected.
| Approach | What your agent gets | A good fit when |
|---|---|---|
| API key plus the REST docs | Works with any agent, but it learns the API from the docs each session | You need one narrow call from a script |
| A hyperscaler’s agent toolkit | One managed MCP server and skills across the provider’s service catalog, with access set by the provider’s credentials | Your workload already lives in that provider’s account |
| A neocloud’s infrastructure MCP | Monitoring and troubleshooting data for workloads that provider runs | You are debugging a job on that provider’s cluster |
| Our skills and the hosted MCP server | Playbooks for AI jobs: pick hardware, deploy a Pod or Serverless endpoint, repair a ComfyUI workflow, audit costs. Every tool carries read-only and destructive labels. | The job is running AI workloads on Runpod and you want the agent to do the setup, deploy and cleanup |
Pick the row that matches where your workload runs.
Install by client
Claude Code and Codex each install one plugin that carries both the skills and the MCP server. ChatGPT uses a plugin listing, and most other agents take two commands.
Claude Code:
claude plugin marketplace add runpod/runpod-plugins-officialclaude plugin install runpod@runpodRun /reload-plugins, then sign in through /mcp.
Codex:
codex plugin marketplace add https://github.com/runpod/runpod-plugins-official.gitcodex plugin add runpod@runpodChatGPT: add the Runpod plugin and sign in, with no terminal needed.
Cursor and most other agents take two commands (Node.js 22.20+ for the first command):
npx skills add runpod/runpod-plugins-official --skill '*' --agent cursor --globalnpx @runpod/mcp-server@latest add --client cursorThe setup docs list the commands and ids for all 17 clients.
Check it works
List my Runpod pods.An empty list is a pass: the connection works and the account has no Pods.
Run it from CI
In CI (continuous integration), a scheduled job or an agent SDK run, an agent can’t complete browser sign-in, so it uses an API key.
- Create a key on the Credentials page and store it as a secret named
RUNPOD_API_KEY. Never print it or write it to a file. - Install the skills as above, skipping
/reload-pluginsbecause the plugin loads the next time the agent starts. - Connect the hosted MCP server with the key as a bearer header. For Claude Code:
claude mcp add --transport http runpod -s user https://mcp.runpod.io/ --header "Authorization: Bearer $RUNPOD_API_KEY"- In headless Claude Code (
claude -p), allow the skill with--allowedTools "Skill(runpod:runpod-mcp)", and add the MCP tools your task needs. Claude Code refuses anything you don’t allow, so a destructive tool such asdelete-podstays blocked unless you list it.
For other clients, add the same Authorization header to the server entry, as the setup docs show.
Troubleshooting
If the Pod list is empty but you expect Pods, the sign-in went to a different Runpod account. Sign out of the Runpod MCP server, sign in with the right account and list Pods again.
Register only one MCP URL. mcp.getrunpod.io is the same server under its old name, and each address has its own sign-in.
Next steps
Want to try a skill on a real job? Ask your agent something like “Audit my Runpod costs and list anything I could stop.” The skills docs list the rest.
Curious how the skills and MCP server are built and tested? Read How Runpod’s agent skills and MCP server work together.
Want to define Runpod in Terraform? Read Manage Runpod as code with Terraform provider 2.0.
Want everything else we shipped this month? Read From setup to deployment: what’s new in Runpod for Agents.
