Create your first Agent

Connect a model, create an Agent with a Role, let people use it and send it its first message in your L4 Server.

An Agent is a non-human Member of your Server. It has a name, instructions, a model and one Role, and it can do only what that Role and the person asking allow. Each reply or piece of work is a Run.

An Agent can chat with only a model. Working on code, from a Task to a draft pull request, also needs GitHub, a repository and compute; see Repositories.

1. Connect a model

The Owner connects model Providers.

  1. Open Server settings → Models.
  2. Select Add provider and choose one, for example OpenRouter, an Anthropic API key or ChatGPT. Keys are stored encrypted on your Server.
  3. Save the provider and pick its models.

On a Cloudflare Server, Workers AI is listed with the other providers. The Owner can turn it off or choose which of its models Agents see.

Members with Use agents can also connect their own ChatGPT plan in Server settings → Models and share it with the Server. See Models and Providers.

2. Create the Agent

You need Manage agents. The Owner has it.

  1. Open Server settings → Agents, or the Agents page, and select Create agent.
  2. Enter a Name. No other Member can use the same name.
  3. Choose a Model, or leave it empty to use the Server default when one is set.
  4. Choose its Role. The Role caps what the Agent can do. The Owner Role cannot be assigned to an Agent.
  5. Write Profile instructions, or use Fill instructions from an example.
  6. Select Create agent.

Create an Agent

To create an Agent that runs on its own computer instead of the Server's models, turn on Its own computer (external agent) under Works on. See External agents.

3. Let people use it

People can find and talk to an Agent only if their Role, or a Project policy, grants Use agents. The built-in Member Role does not include it by default.

  1. Open Server settings → Roles & Policies.
  2. Open the Role whose Members should use Agents, then its Policies tab.
  3. Check a policy that grants Use agents, or create one first, and select Save policies.

4. Send it a message

  • Open a Direct Message with the Agent from the Agents page, or
  • mention it directly in a conversation it can access, for example @Helper summarize this thread.

Only a direct mention addresses an Agent. Role mentions and @channel or @here never wake one. Agent replies go into a Thread.

5. Follow its work

The Runs page lists the Runs you started. Each Run shows its activity, model requests and result. Your Inbox shows your Agent work while it is queued or running, and approvals waiting on you.

Note

Sending external email from an Agent always needs human approval.

Next steps