Get Started

Agent runtimes

Which agent CLIs Froots runs, how they sign in, and how to pick one per agent.

Froots doesn't ask for model keys. Every agent runs on an agent CLI (Froots calls them runtimes, or harnesses), and each CLI signs in on its own: your Claude subscription, codex login, and so on. Froots launches the CLI and drives it. Change your mind any time, per agent.

The runtimes

Four ship built in and always show in Settings → Agents:

  • Claude Code — install it (npm install -g @anthropic-ai/claude-code), run claude once, and sign in.
  • Codex — install it (npm install -g @openai/codex) and run codex login.
  • Gemini CLI — runs through gemini --acp. Sign in through the CLI itself.
  • Goose — runs through goose acp.

More are in the Add runtimes catalog: Amp, Cursor, Grok Build, Hermes Agent (hermes acp; run hermes model once to pick a provider), Kilo Code, Kimi Code, Oh My Pi, and OpenCode. OpenClaw is listed too, but it doesn't work yet.

A runtime that isn't installed stays out of the model pickers. It still shows in Add runtimes, with its setup guide. On Mac and Linux, runtimes with a one-line installer also get an Install button that shows the exact command before it runs. Once the CLI is on your PATH, it shows up everywhere.

Running something else? Add runtimes → Custom harness takes a command, its arguments, and any environment variables. It works with anything that speaks the Agent Client Protocol (ACP).

Signing in

Froots never holds your model login. The CLI does. If a runtime isn't signed in, your message fails with a notice like:

Claude Code isn't signed in. Open a terminal, run claude and sign in, then send your message again.

Do that, then hit Retry.

Want an API key or a local model instead of a subscription? Set it up in the runtime itself. OpenCode, Kilo Code, Goose, and Hermes Agent each let you choose a provider, and Froots uses whatever the CLI is configured for.

Choosing a runtime and model

  • During setup — the "Choose your default harness" step sets the runtime your first agent uses.
  • In a chat — in a one-on-one chat with your agent, the picker in the top-right switches runtime and model for that agent.
  • Per agent — right-click an agent, choose Edit agent, and open Models. You can set the runtime, the model, and how hard it thinks. The same tab holds Where this agent runs (see below).

Models come live from each runtime, so new ones appear as soon as the CLI knows about them.

Where this agent runs

If you use Froots on more than one computer, each agent answers from one place:

  • Follow me — it answers on whichever device you message it from. Files it makes stay on that device.
  • A specific device — it always answers from that machine. If that machine is off, another one of yours stands in until it's back.

Your other devices show up in the list once they're running a current version of Froots.

Mixing runtimes

Nothing forces one runtime on everything. A common setup is Claude Code for long coding runs, Codex for a second opinion, and an open-source runtime pointed at a local model for private work. Each agent carries its own runtime and model.