Docs

Router engines

When you press the Talk hotkey and say or type what you need, Orbit's router works out which agent should get it, or whether to start a new one, and writes a short receipt back to you. This page is about the model behind that step.

Default: Claude Code on your primary Mac

Out of the box the router runs claude -p on your primary Mac, using the Claude Code login and plan you already have. It runs headless with no tools and saves nothing. Each request is one short call: the router sees your sentence, the list of agents with a few recent lines from each, your Notes, and the screenshot if you pointed at something. It answers with a destination and a one-line receipt.

Nothing is sent anywhere except from your Mac to Anthropic, the same way Claude Code itself works. Orbit's relay and servers are not involved.

Option: an Anthropic API key

If Claude Code is not installed on the Mac, or you would rather bill routing separately, add an Anthropic API key in Settings → Router. The router then calls the API directly from your Mac with that key. You can pick the model and effort there too.

Without either

The router still works with rules: name an agent ("tell claude-7 to…", "claude-7: …") or a project directory and the request goes there. Recaps of finished work use the agent's own session title and final message.

What the router can do

• Send your request to the agent it is clearly about.
• Answer an agent's open question when your request is an answer.
• Start a new agent in a project directory you have listed, when nobody is on it.
• Answer you directly for questions like "what is everyone doing?"
• Tell you when the agent it chose finishes.

Other engines

Codex, local models and other providers as router engines are planned. The router is a small, well-defined interface, and we will add engines as they prove reliable at this job. If you need one in particular, tell us.