Claude, Codex and Pi
assist runs agents in a coding harness: Claude Code, Claude (SDK), the Codex CLI or pi. Claude is the default, and the others work wherever they’re installed.
Setting up Codex and pi
Section titled “Setting up Codex and pi”Install the harness, then run assist sync. It installs assist’s slash commands into each harness it finds on your PATH, and the hooks that let read-only commands run without asking.
Claude (SDK)
Section titled “Claude (SDK)”Claude (SDK) runs your installed, signed-in Claude Code without a terminal: the session shows its transcript and a composer instead. It always runs in auto permission mode, so it never stops to ask permission. Pick it as the harness when starting a session, or make it the default with assist config set harness.engine claude-sdk. Signed in with a Claude plan, it uses the same plan usage limits as Claude Code.
Choosing a harness
Section titled “Choosing a harness”Actions that start an agent offer with Codex or with pi wherever that harness can run them. On the CLI, every command that starts an agent takes --harness <claude|claude-sdk|codex|pi>.
To use another harness by default, set harness.engine:
assist config set harness.engine codexTo pick a harness, and a model on it, for one type of session, set harness.defaults.<type>.harness and harness.defaults.<type>.model, where <type> is prompt, draft, bug or design in the New session dialog, or run for backlog runs. A type with no harness set uses harness.engine, and one with no model uses the harness’s own default:
assist config set harness.defaults.draft.harness codexassist config set harness.defaults.run.model opusHiding a harness
Section titled “Hiding a harness”The with Codex and with pi menu entries show whenever that harness is installed and can run that action. To hide them, turn the harness’s key off:
assist config set harness.exposeCodexActions falseassist config set harness.exposePiActions falseTo run Codex on a model served by your own proxy, see LiteLLM.
