Skip to docs content
Browse documentation
On this page

Connect your agent

The three-step Get started flow for people: connect the agent you already use over MCP, paste one prompt so that agent makes something useful in your account, then connect a second agent from another ecosystem and watch it reuse the same memory and package. Includes what approving access means, verified-email requirements, and which agents suit building versus using.

Minion is an MCP server. You use it from Claude, ChatGPT, Cursor, Codex, Claude Code, or any other agent that supports MCP — not from a separate Minion chat app. Getting started is three steps, and the in-app Get started page at /app/onboarding walks you through each one.

Before you start

  • A Minion account and a team. Signing in puts you in a team (organization); everything your agents save belongs to it.
  • One agent you already use. You do not need every app on day one. A second agent is Step 3, and it is worth it precisely because the first one has already furnished the home.
  • A verified email if you want your agent to send you mail. Sending needs the address on your account to be verified.

Step 1 — Connect one agent

  1. Open Get started (/app/onboarding). It shows the MCP URL for your deployment. On app.minion.codes that is https://app.minion.codes/mcp.
  2. Add that URL to your AI app. Claude (claude.ai or Claude Desktop): Settings, Connectors, Add custom connector. ChatGPT: Settings, Apps and Connectors, turn on Developer mode, then Create. In both, paste the URL, choose OAuth if asked, then sign in to Minion and approve.
  3. For Claude Code, Cursor, Codex and other headless clients, create an API key at /app/api-keys. The page shows the exact command or config to paste, for example claude mcp add --transport http minion <url> --header "Authorization: Bearer <key>". A key acts as you, so treat it like a password.

Approving gives that agent access to this Minion team, acting as you. Minion does not control or supervise what your agent does with the access you grant. Connect agents you trust, and revoke one from the Connections page (/app/connections) when you stop using it.

Some apps bind MCP tools when a conversation starts. If Minion's tools do not appear right after authorizing (Claude Desktop is the usual case), start a new chat before the first task.

Which agent should go first?

Using packages that already exist works well from chat apps: Claude, ChatGPT and similar. Creating or editing packages is smoother from a coding agent — Claude Code, Cursor, Codex — because those hosts edit files and iterate on code easily. Either kind can be first; the home is the same.

Step 2 — Say hello

Once the connection works, Get started shows a short prompt to paste into the agent you just connected. The same prompt is available over MCP as the onboarding_setup prompt. The agent explains what Minion can do for you, calls search with an empty query to show the domains, picks one small thing you could automate, smoke-tests it with execute, and saves that working code as a package. The page ticks "Minion heard from your app" when the agent's first search arrives.

This is the moment Minion stops being a login and becomes yours. The agent should not tour every surface or wire up integrations you did not ask for. If it starts talking about connecting APIs before it has made anything, point it back at the first-run briefing.

Two optional playbooks sit next to that step: First build turns one ad hoc execute into a package you own, and Email and memories uses the team inbox to save durable memories about you.

Step 3 — Connect a second agent

Step 3 is where Minion earns the name. Add the same MCP URL to a different app (say ChatGPT if you started with Claude) and ask it what Minion knows. Packages, secrets and memories are shared across every app you connect, so the new agent finds the memory or package you just made — in a different product, with no setup repeated. The page ticks "two apps share your Minion" when both have connected.

Where to go next