MCP is the standard way AI apps use outside tools. The onpod CLI includes an MCP server, so you only need to connect the CLI you installed to your AI app. It works even in Claude Desktop, which has no terminal, and login, features and pricing are the same as the CLI.
“Put this folder on onpod”
“Show me how my apps are doing”
“Why won't the blog app start? Check the logs”
👉 For the no-install “connector” that connects with just a URL (claude.ai · Claude desktop · classes), see Deploy by talking to Claude. This page is about connecting the onpod CLI on your own computer (servers and databases included).
How to connect
- ① Install the CLI and log in once — with the key from "API keys" in the console:
onpod login --token … - ② Add the settings below to your AI app
- ③ Restart the AI app and the onpod tools appear
Claude Code · Codex · Gemini CLI (one line in the terminal)
claude mcp add onpod -- onpod mcp codex mcp add onpod -- onpod mcp gemini mcp add onpod onpod mcp
Claude Desktop · Cursor · Windsurf
Claude Desktop: Settings › Developer › Edit Config. Cursor: ~/.cursor/mcp.json. Windsurf: ~/.codeium/windsurf/mcp_config.json.
{
"mcpServers": {
"onpod": { "command": "onpod", "args": ["mcp"] }
}
}VS Code
Put it in the project's .vscode/mcp.json.
{
"servers": {
"onpod": { "type": "stdio", "command": "onpod", "args": ["mcp"] }
}
}Without installing the CLI (npx)
All you need is Node.js. Instead of logging in, put your API key in the env of the settings.
{
"mcpServers": {
"onpod": {
"command": "npx",
"args": ["-y", "onpod", "mcp"],
"env": { "ONPOD_TOKEN": "opk_..." }
}
}
}If the AI app says it cannot find onpod, run onpod mcp config in a terminal. It prints the settings for each app with the exact path of onpod on this computer (desktop apps search a different path than your terminal).
Tools the AI uses
| Tool | What it does | Type |
|---|---|---|
| manual | How onpod works — reads only the topic it needs | Read |
| whoami | Shows the connected account and company | Read |
| apps | Your apps with state and URL (filter by name) | Read |
| app_status | Detailed status of one app and why it failed | Read |
| logs | Last N lines of app or server logs | Read |
| deploy_plan | Looks at a folder and proposes a deploy with the monthly price (creates nothing) | Read |
| deploy | Deploys as proposed and checks the public URL | Change |
| exec | Runs one command inside a server | Change |
| onpod | Every other CLI command (DB, domains, GPU, file storage…) | Change |
| job | Waits for or cancels work that takes minutes | Read |
"Change" tools ask for your confirmation in the AI app before they run. Deploys still go propose → confirm → run, and a deploy that takes minutes is not cut off by the AI app's time limit: it returns a job number (job_id) first and finishes in the background — the AI picks up the result with that number.
Safeguards
- Login is never done over MCP — use CLI login or the env in the settings so your key never lands in the chat history
- Payment cards, auto-recharge settings and operator commands are not available (creating a top-up link is)
- Secret folders like ~/.ssh or ~/.aws and uploading your whole home folder are refused. Files downloaded from a pod are saved only inside the working folder you choose
- To allow reading only, set args to ["mcp", "--read-only"] — the deploy and command tools disappear entirely
- For extra safety, create a separate API key with only the permissions and expiry you need and use it for MCP
Pricing
MCP and the CLI are free. You pay only for what you run — apps, databases, GPUs — at the usual prices.
Using it at a company (organization account)
Log in with an organization API key (or put it in the env) and everything the AI creates is a company asset from the start, paid from the company wallet. Requires CLI 0.7.137 or later.