Build your own agent on Argorant

If you are writing the runtime yourself, you have three ways in. Pick by how your agent executes, not by preference.

• MCP at https://mcp.argorant.com/mcp: best when your framework already speaks MCP and you want OAuth sign-in instead of key handling.
• REST at https://argorant.com/api/mcp/... with a bearer key: best for servers, cron jobs and anything headless.
• CLI (npx argorant): best when the agent can shell out and you want files on disk plus machine-readable exit codes.

A loop that behaves

1. Call GET /api/mcp/account once at startup to confirm the key, its scopes and the daily quota.
2. Translate the user's request into filters and call GET /api/mcp/people/count. Free, so iterate here.
3. Show a redacted preview with GET /api/mcp/people/preview and let the human confirm the shape.
4. Only then reveal or create an export, with an explicit row limit.
5. Handle the four failure codes deliberately: 401, 402, 403 and 429 each mean something different.

Use --json on every CLI call inside an agent, or read the JSON bodies straight from REST. Both give stable field names. The AI reference at argorant.com/ai is written to be pasted into an agent's context.