Glossary
what is a harness router?
TL;DR
A harness router is a service that lets software send a task to a ready-made coding agent, like Claude Code, Codex or Hermes, through one API. It picks the agent, runs it headless in an isolated sandbox, and streams back progress, results and files. It's the agent-level version of a model router: instead of choosing which model answers, it chooses which whole agent does the work. It's built for products that embed agents, not for a developer running agents by hand.
If you build a product that needs a coding agent inside it, you can write your own agent, or you can call an existing one. A harness router is the second option, made easy: one API in front of many agents.
How it works
Your app sends a task; the router runs an agent in a sandbox and sends back the result.
Your product sends a task over HTTP and names an agent, or lets the router pick. The router starts that agent headless inside a sandbox, passes it the task and any files, and streams back what the agent is doing. When the agent finishes, you get its output and the files it changed.
HarnessRouter is the clearest example. It pitches itself as "OpenRouter for agent harnesses", runs Codex, Claude Code, Hermes and about 17 other agent CLIs, and publishes an open spec for the API called the Unified Harness Protocol. It also offers a free self-hosted edition as a Docker image.
Harness router vs model router
A model router picks who answers a prompt. A harness router picks who does a whole job.
A model router, like OpenRouter or OpenCodex, sits between an agent and the AI providers and decides which model answers each request. The agent stays the same.
A harness router sits one level higher. It chooses the whole agent, with its own tools, loop and habits, and runs it to the end of a task.
| Routes to | |
|---|---|
| Model router | A model, for each request from an agent you already run |
| Harness router | A whole agent, for a whole task, run in a sandbox |
| Agent manager | Nothing: you choose and steer agents yourself, on your own machine |
Who it's for
Teams building agent features into a product, not developers running agents by hand.
A harness router suits a product team that wants an "agent backend": a feature that writes code, fixes a bug or edits documents for its users, without building an agent from scratch.
If you're a developer running agents on your own code, you want an agent manager instead. Pane, for example, runs Claude Code, Codex and others in real terminals on your machine, each on its own git worktree, and you review the result yourself.
Related
Sources
- Corrections
- If anything here is wrong or out of date, tell us and we'll fix it and update the date at the top. Open an issue on GitHub
- Sources
frequently asked questions
By Parsa Khazaeepoul, co-founder of Pane. .