Skip to main content
A runner is the AI engine that processes your prompts. RemoteAgent.CHAT acts as the relay — the runner is what actually reads your code, reasons about it, and produces output. You choose a runner once during remoteagent init, and you can change it at any time by re-running init.

​
Comparison table

If you have ANTHROPIC_API_KEY set in your shell environment, the claude binary will use it regardless of any other authentication — it overrides both the OAuth token and claude login, even with an active Pro/Max subscription. Unset it if you want to authenticate via subscription:
See claude-code authentication for the full details.

​
How to select a runner

Run remoteagent init and the wizard will show an interactive list — use the arrow keys to pick a runner, then press Enter:
If you already know which runner you want, you can skip the prompt by passing --runner:

​
Switching runners

To change the runner for an existing agent, run remoteagent init again in the same project directory. The wizard will detect your existing configuration and ask what you want to do — choose “Update runner — keep existing pairing” to change only the runner without re-pairing. After the wizard completes, start the agent:

​
Runner architecture

All runners implement the same interface internally. When the agent receives a command, it calls the runner with the prompt and streams chunks back as they are produced. The runner is responsible for:
  1. Receiving the prompt string
  2. Invoking the AI model (in-process, via subprocess, or via HTTP)
  3. Yielding output chunks as they arrive
  4. Signaling completion or error
This uniform interface means the Telegram experience is identical regardless of which runner you use — output always streams in real time.
Runners marked Beta are functional but may have edge cases in output parsing or error handling. If you encounter issues, please open an issue on the GitHub repository.