Skip to main content
The shortest path to a live agent uses a workspace API key and three HTTP calls. You need no keypair, no JWKS document, no connection, and no endpoint of your own — those arrive later, when the agent needs to call your backend.
1

Get an API key

In the dashboard, open Developers and copy a secret key (sk_…). Keep it server-side.
2

Publish an agent

Follow Your first agent. You need its public id, which looks like agent_4f3c….An agent with no patches wired still works — every agent can ask questions and hand off out of the box.
3

Make the three calls below

1. Open a thread

A thread is one conversation, and it belongs to a subject — an opaque id you choose for whoever the agent is acting for.

2. Say something

202. The turn is accepted, not finished — a run is the unit of work, and it is still going.

3. Read the reply

Poll until terminal is true. That is your agent working.
A run can also stop and wait for you — with a question, or with a write that needs approval. Both show as awaiting_input, and both are resumed differently. See Runs and states.

Where to go next

You now have a conversation. What you probably want next is an agent that does something in your system, and a way to put it in front of real users.

Choose an integration

Relay from your backend, or direct from the browser.

Connect your backend

Register a host so the agent can call your endpoints.

Patches

Wrap your endpoints as capabilities the agent can use.

Subjects

Get the isolation boundary right before you have real users.