Overview

A harness is the agent runtime that drives the loop — the CLI or service that reads your prompt, calls tools, and edits files. Each one already speaks some model API; the move is always the same: point it at the local proxy at http://127.0.0.1:4356 instead of the vendor’s, and address models by their provider/model id. From there the same harness can run on Anthropic, OpenAI, Google, or an open model — with provider selection and fallback underneath.

For a one-shot launch that leaves the harness’s normal config alone, run Grok through routeplane spawn.

Harness vs. model source

Two of these share a name with a model source, and it’s worth keeping them straight:

You want to…	Use
Run the Claude Code CLI on any model	Claude Code (harness)
Spend your Claude plan as tokens	Claude subscription (model)
Run the Codex CLI on any model	Codex (harness)
Spend your ChatGPT plan as tokens	Codex subscription (model)

A harness is what runs; a model source is where the tokens come from. You can combine them freely — e.g. drive Claude Code (harness) against your ChatGPT plan (Codex subscription), or against a local Ollama model.

One-shot launch with routeplane spawn

If you do not want to edit a harness config, routeplane spawn injects Routeplane routing into a child process. It supports Claude Code, Codex, Kimi, OpenCode, and Grok. For example:

routeplane spawn -a grok -m openai/gpt-5.4 -- -p "review this diff"

The chosen model is a Routeplane provider/model id. Arguments after -- go to Grok. The Grok path uses a temporary isolated home so its cached provider URL cannot bypass Routeplane, and it removes that home when the child exits. See the CLI reference for every supported harness and the fail-closed flags.

New to Routeplane?

Start with the Quick Start to get the proxy running, then come back and point your harness at it.