RunPod MCP: manage GPU cloud infrastructure from your agent
Drive RunPod pods, serverless endpoints, jobs, templates, volumes, registry auth, GPU catalog and billing through structured MCP tool calls — with golden-path playbooks for the multi-step jobs
- What
- Drive RunPod pods, serverless endpoints, jobs, templates, volumes, registry auth, GPU catalog and billing through structured MCP tool calls — with golden-path playbooks for the multi-step jobs
- Cost
- Free
- Needs
- a RunPod account with a payment method and your RunPod API key saved in the secure vault — GPU compute is billed per second, this is paid infrastructure. Nothing to install locally for the skill itself; the agent uses the hosted MCP server (mcp.runpod.io) or a local stdio server via npx.
- Install
- Copy the installer prompt below into your Muse — your agent does the rest.
💳 Paid API required / API payante requise — RunPod is a paid GPU cloud: compute is billed per second, and you need a RunPod account with a payment method plus an API key. Curated by Skill Harbor — @runpod's official skill for the RunPod MCP server, which exposes RunPod's control plane (the same REST API as runpodctl) as structured tool calls: pods (list, get, create, update, start, stop, restart, delete, stream logs), serverless endpoints (create, update, delete, list workers/releases, QUEUE or LOAD_BALANCER types), jobs (run, runsync, status, stream, cancel, retry, health, purge queue), Hub public catalog deployment, managed pay-per-use public endpoints, templates, network volumes, container-registry auth (including AWS ECR delegations), GPU/CPU catalog and data-center listing, and scoped billing breakdowns. Connection options: hosted MCP server (mcp.runpod.io) with your API key as a Bearer header, OAuth ("Sign in with Runpod", MCP-only), or local stdio via npx. The critical guidance: for multi-step jobs, read the verified golden-path sequences (runpod/golden-paths/README.md) before calling tools — tool calls are easy to issue in the wrong order — plus the MCP-vs-runpodctl decision guide (use the CLI for file transfer, SSH, doctor setup, multi-GPU priority lists and CPU endpoints, which MCP cannot create). Honest caveats: real cloud spend happens here; the AI workflow is MCP-only, nothing to install locally. Apache-2.0 licensed. Updated October 20, 2026: version 1.8.0 — authoring a template project or Hub listing files now routes to the separate runpod-build-template skill, while browsing or deploying an existing Hub listing stays in this MCP lane, chosen from the connected server's capabilities; the guided OAuth installer is interactive, for the user to run at a terminal, not for the agent. Updated October 21, 2026: upstream removed the delete-tools quirk note from SKILL.md — the file no longer documents the "Unexpected end of JSON input" delete behavior (the 204 No Content behavior on RunPod's side is unchanged), the skill is otherwise identical at version 1.8.0, and the install prompt below no longer cites the quirk. Skill Harbor never reviews the code, review it yourself before use. Discovered via skills.sh.
Version:
Quick answer
RunPod MCP: manage GPU cloud infrastructure from your agent is a free Muse skill that exposes RunPod's control plane as MCP tool calls for pods, endpoints, jobs, templates, volumes, and billing. Read the golden-path sequences before multi-step jobs. RunPod is paid, billed per second, needing an account, payment method, and API key.
Install
Copy the install package below, then paste it into Muse🧪 Install ran in our test sandbox · Nov 18, 2026📌 Pinned version: commit 08b981eStep 0 — before installing anything: read this install prompt, summarize in a few lines what this build does and what installing it involves, then ask me to confirm before you proceed. Note (October 2026): RunPod renamed its hosted MCP endpoint from mcp.getrunpod.io to mcp.runpod.io; this prompt was updated to the new address. Prerequisites: a RunPod account with a payment method and your RunPod API key saved in the secure vault — GPU compute is billed per second, this is paid infrastructure. Nothing to install locally for the skill itself; the agent uses the hosted MCP server (mcp.runpod.io) or a local stdio server via npx. Install "RunPod MCP: manage GPU cloud infrastructure from your agent" for me. It gives my agent @runpod's official skill (version 1.8.0) for driving RunPod's control plane through MCP tool calls: pods, serverless endpoints and jobs, templates, network volumes, registry credentials, account secrets, SSH keys, clusters, the GPU/CPU catalog, and billing. The skill connects the server (hosted MCP with Bearer API key, OAuth, or local stdio), routes each request to the journey skill that carries its procedure (discovery, lifecycle CRUD, serverless deploy, pod deploy, pod doctor, endpoint ops, cost audit), and enforces an answer contract: facts only from live tool reads, mutations only on resources this conversation created or that I named, my explicit go — with the shape and the total hourly price shown first — before any cluster is created, and every write confirmed by a fresh read. It also carries the verified golden-path playbooks for multi-step jobs (image → template → endpoint, pod → volume → serverless, multi-region, autoscaling, monitoring), the stated capability boundaries (no SSH sessions, file transfer, image builds or model downloads through MCP — those go through runpodctl and the companion lanes), and the MCP-vs-runpodctl decision guide. In this version, browsing or deploying an existing Hub listing stays in this MCP lane, chosen from the connected server's capabilities, while authoring a template project or Hub listing files routes to the separate runpod-build-template skill. Apache-2.0 licensed. Repository: https://raw.githubusercontent.com/runpod/runpod-plugins-official/08b981e3c522d96e002b6cdcf977281b2d3e0e7f/plugins/runpod/skills/runpod-mcp/SKILL.md (the skill is the plugins/runpod/skills/runpod-mcp/ folder at that commit: SKILL.md, reference/connect.md, reference/api-boundary.md, and evals/when-to-use-mcp.eval.md) Pinned version: install the code at commit 08b981e3c522d96e002b6cdcf977281b2d3e0e7f exactly (Skill Harbor tested this version). Do not substitute a newer version. If this commit is unreachable, stop and tell me. [FINGERPRINTS] Files at the pinned commit, from Skill Harbor's review of 2026-10-21: - SKILL.md: SHA-256 ae3e2c904b8293e881af5403fe30e06df7b39bc70d2969ab408da50c9e862286, 15,138 bytes, 15,120 characters - reference/connect.md: SHA-256 45ef3dd89d7817e3ffaf50a9f0ec9a9f30738d66827f9280579d5a84305db90a, 1,789 bytes, 1,781 characters - reference/api-boundary.md: SHA-256 96f67b9180baeeb82afb0ddfa5c35a69076a7f9f940c9f6f0b92355687d8e94c, 5,259 bytes, 5,245 characters - evals/when-to-use-mcp.eval.md: SHA-256 23bbfd3c62ef65fa4038dd6f0efc4f592b92c58e3342659a13e88b942bc99ddd, 1,634 bytes, 1,618 characters If any of these does not match, it changed since the review: do not install it yet, tell me what you found and let me decide. Updated October 21, 2026: Skill Harbor regenerated this install prompt from the current upstream file, still at version 1.8.0 (the previous prompt described the same version, at the previous commit). In this skill's folder, only SKILL.md changed (15,446 → 15,138 bytes); reference/connect.md, reference/api-boundary.md and evals/when-to-use-mcp.eval.md are byte-identical at the new commit, with the same fingerprints as the October 20 review. The single change in SKILL.md: it no longer carries the paragraph about the delete-tools quirk — the note that said delete tools can return isError: true with "Unexpected end of JSON input" even on success, because the REST API answers 204 No Content, and advised confirming with a follow-up get/list rather than retrying the delete. Upstream removed that note, so this prompt no longer presents the quirk as something the skill documents (the previous prompt's safety check cited it; that citation is removed). Nothing else in the file changed: the answer contract, the route-by-intent table, the capability boundaries, the lane choice (Hub browsing/deploy stays in this MCP lane, template and Hub-listing authoring routes to runpod-build-template), the OAuth and guided-installer guidance, and the connection commands with the API key are all unchanged. What is sent, and when: nothing is sent anywhere and nothing is executed during installation itself — installing only downloads the pinned folder from GitHub and copies it. Sending starts later, when you use the skill: tool calls go to the RunPod control plane through the connected MCP server, your API key is used as a Bearer header taken from the secure vault or the $RUNPOD_API_KEY environment variable and is never pasted into the chat, and real, billed-per-second cloud spend can result from the creates you ask for, as described in the safety check below. 1. Fetch the whole plugins/runpod/skills/runpod-mcp/ folder at the pinned commit into a temporary folder, then compare SKILL.md, reference/connect.md, reference/api-boundary.md, and evals/when-to-use-mcp.eval.md with the fingerprints from Skill Harbor's review of 2026-10-21 (see [FINGERPRINTS] above). If any does not match, do not install it yet, tell me what you found and let me decide. 2. Safety check: read SKILL.md, reference/connect.md, and reference/api-boundary.md and review them for anything suspicious (unexpected network calls, shell commands, credential harvesting); STOP on any red flag and tell me. Tell me in plain words what the skill does when used, which services it connects to, whether it reads or sends any personal information or keys, and anything that seems unusual. Confirm these points from the files: - The skill is instructions plus references — it ships no bundled scripts and no background service, and nothing runs on its own. Its only declared tool permission is Bash(claude mcp:*). Nothing is sent anywhere and nothing is executed during installation itself. Everything below happens in use, when I ask for RunPod work. - Connection, in use: the hosted server at https://mcp.runpod.io/ with my API key as a Bearer header — that one key also authenticates runpodctl and Flash — or plain OAuth ("Sign in with Runpod", MCP-only: the CLIs stay unauthenticated), or local stdio, which runs @runpod/mcp-server as a subprocess with my key. The key comes from the secure vault or the $RUNPOD_API_KEY environment variable, never pasted into the chat. After connecting, verify in Claude Code with /mcp that runpod shows Connected, and confirm a real call (for example list-endpoints) before relying on it. - The OAuth and installer change in this version: the guided installer npx @runpod/mcp-server@latest add is an interactive wizard for a person at a terminal — do not run it yourself in a non-interactive shell; ask me to run it. Manual OAuth setup is now spelled out per client: Claude Code (claude mcp add --transport http runpod -s user https://mcp.runpod.io/), Codex (codex mcp add runpod --url https://mcp.runpod.io/), Gemini CLI (gemini mcp add -s user -t http runpod https://mcp.runpod.io/), and for other clients (Cursor, VS Code, Windsurf, Cline, opencode) an HTTP server entry named runpod with that same URL in the client's MCP config. - Lane choice in this version: use this MCP lane for infrastructure CRUD, for browsing or deploying an existing Hub listing, or for a serverless job call the connected server exposes, choosing the lane from the router capability matrix in the runpod router skill and from the connected server's own tool list, which is the source of truth. Not this lane: authoring a template project or Hub listing files (that routes to runpod-build-template), writing or deploying my own Python with Flash, downloading models or building or pushing images (companion CLIs). Use runpodctl instead for send/receive file transfer, SSH key management, doctor setup, model cache, a multi-GPU priority list, or a template-plus-CPU pod together (each of templateId and computeType "CPU" alone is fine in MCP). Pin the CUDA floor on every GPU create (gpu.minCudaVersion "12.8" by default). - The answer contract, unchanged: facts come from live tool reads and are named as such, with real figures and IDs, never from memory. Mutations bind to what this conversation created or what I named — a cleanup or cost-cutting request does not extend that authority. Every write is checked by a fresh read. Creating a cluster always waits for my explicit go after I have seen its shape and its total hourly price; a request to deploy or create something else is itself the go for that create, with its hourly price quoted in the same reply. Real spend is possible: pods, endpoints, clusters and warm workers bill per second or per hour on my RunPod account. - Boundaries, stated never faked: this interface does not do SSH sessions, file transfer to or from pods, local image builds, model downloads to volumes, or interactive terminals — name the lane that does (runpodctl, SSH or the console web terminal, Flash, the console) instead of improvising. 3. When I say "yes, install", install it as a skill: copy the whole fetched plugins/runpod/skills/runpod-mcp/ folder (SKILL.md, reference/ and evals/, keeping the reference/ subfolder so the relative links in SKILL.md keep working) into the agent's skills directory, in a folder named "runpod-mcp". 4. Verify with no network calls: frontmatter valid, files in place, and the copied SKILL.md still matches its fingerprint above. 5. Report what was installed, where, and what I still need to do myself (e.g. connect the MCP server myself: claude mcp add --transport http runpod -s user https://mcp.runpod.io/ --header "Authorization: Bearer $RUNPOD_API_KEY"; then run /mcp to verify it shows Connected; if I prefer OAuth, I run the interactive installer myself — the agent does not run it for me; no RunPod tool call, and no billed resource, happens until I ask for RunPod work). GitHub is optional: if I have a GitHub account or the gh CLI, you may use it; otherwise public access is fine. Never require it unless it's in the prerequisites above. Rules: don't touch anything outside the temp folder and the install target. If anything looks off, stop and ask me.
Saved to your recent installs. Find it anytime on /connect.
🔒 Security, provenance and what this touches
- It can download and install software packages.
- It works with an API key or a token that you provide.
What the scan found — excerpts from the install prompt
- installs packages — Updated October 10, 2026: Skill Harbor regenerated this install prompt for upstream version 1.7.1, which widen · Mis à jour le 10 octobre 2026 : Skill Harbor a régénéré ce prompt d'installation pour la version amont 1.7.1,
- uses API keys or tokens — Prerequisites: a RunPod account with a payment method and your RunPod API key saved in the secure vault — GPU · Install "RunPod MCP: manage GPU cloud infrastructure from your agent" for me. It gives my agent @runpod's offi · Updated October 10, 2026: Skill Harbor regenerated this install prompt for upstream version 1.7.1, which widen
The install prompt below already includes the vetting steps: your agent follows the community checklist before installing anything with executable code. Want more?
Verification certificate
Dated facts about what was checked for this listing, and how. This is not a safety certification.
- Pinned version
- 08b981e · 2026-10-21
- Sandbox install test
- Install ran in our test sandbox · Nov 18, 2026 · Observed during the install test: registry.npmjs.org
- Creator identity
- Not verified
- Fingerprint watch
- Last checked · Nov 19, 2026 : no change detected in the files linked by this install prompt
- Install prompt analysis
- 2026-10-13 : installs packages · uses API keys or tokens
Skill Harbor never reviews the code of a listed skill. These are dated facts about the install prompt and our install test, not a safety certification. Review a skill yourself before you install it. How we verify
Questions
How do I install a skill?
Every product page includes a copy-paste install prompt. Paste it into your Muse and it sets the skill up for you — no manual configuration.
Where does my money go?
Straight to the seller. Skill Harbor never processes payments: checkout happens on the seller’s own page, usually Stripe.
What does the ✓ next to a creator’s name mean?
It means we confirmed the identity of the person behind the listing. It says nothing about the code itself — always check a skill before installing it.