---
name: runpod
description: Use when discovering Runpod API documentation, OpenAPI schema, Serverless endpoint request patterns, and GPU infrastructure workflows.
---

# Runpod Agent Skill

Runpod is the AI Developer Cloud for deploying and managing AI workloads across Pods, Serverless endpoints, Clusters, public endpoints, and related infrastructure resources.

Use this skill when an agent needs to:

- Discover the Runpod REST API, OpenAPI schema, or API authentication model.
- Create, inspect, update, or delete Pods, Serverless endpoints, templates, network volumes, or container registry auths.
- Send requests to Runpod Serverless endpoints with /run, /runsync, /status, /stream, /cancel, or /retry.
- Choose the right Runpod workflow for development, inference, training, batch jobs, or infrastructure automation.
- Find Runpod documentation, examples, or MCP-based docs search.

Primary resources:

- Documentation: https://docs.runpod.io/
- REST API overview: https://docs.runpod.io/api-reference/overview
- OpenAPI schema: https://docs.runpod.io/api-reference/openapi.json
- Serverless request docs: https://docs.runpod.io/serverless/endpoints/send-requests
- Docs MCP server card: https://docs.runpod.io/.well-known/mcp/server-card.json
- Status page: https://uptime.runpod.io/

Authentication:

- REST API calls use `Authorization: Bearer RUNPOD_API_KEY`.
- Never ask users to paste API keys, tokens, or secrets into chat.
- Prefer environment variables or the user's approved local secret manager when running commands.

Common Serverless operations:

| Operation | Method | Purpose |
| --- | --- | --- |
| `/runsync` | POST | Submit a synchronous job and wait for a result. |
| `/run` | POST | Submit an asynchronous job and receive a job ID. |
| `/status/{job_id}` | GET | Check job status and retrieve available output. |
| `/stream/{job_id}` | GET | Stream incremental output when supported. |
| `/cancel/{job_id}` | POST | Cancel a running or queued job. |
| `/retry/{job_id}` | POST | Retry a failed or timed-out job. |
| `/health` | GET | Check endpoint health. |

When unsure, start with the docs MCP server card or the OpenAPI schema, then confirm destructive actions with the user before changing live infrastructure.