Introduction

team-harness is a coordination layer for other coding-agent harnesses. A coordinator LLM receives your task, breaks it into work units, and delegates execution to worker CLIs — Codex, Gemini, Claude Code, Antigravity, OpenCode, pi, and OpenHands — that it spawns as subprocesses. The coordinator never implements; it orchestrates.

It ships as a single command, th (with a team-harness compatibility alias), plus a Python SDK. Point it at an OpenAI-compatible API or a ChatGPT Codex subscription, tell it which workers you have installed, and describe the task.

What it does

You give the coordinator a goal like this:

Tell me what pieces are still missing for the MVP.
 
Create an agent team to do it. They should be responsible for:
- coming up with the analysis using CODEX, CLAUDE and GEMINI
    - each writes its findings into a file in the dedicated directory
- creating the final report
    - read all the analyses and write the final SUMMARY.md

team-harness coordinates the work across Codex, Claude Code, and Gemini — spawning each worker, collecting its output, and driving the plan to completion.

You could get a similar result from Claude Code's built-in agent teams. team-harness differs in two ways: you can plug in any model or provider, and you can tweak the underlying coordinator and worker prompts far more easily.

Why team-harness

• Any model, any provider. The coordinator talks to any OpenAI-compatible endpoint (OpenRouter, OpenAI, local proxies) or a Codex subscription. Each worker keeps its own native auth, or you can route it through OpenRouter too.
• Real CLIs, not reimplementations. Workers are the actual codex, gemini, claude, and agy binaries as subprocesses — you get their real capabilities, sandboxes, and session handling.
• Editable prompts. The coordinator base prompt, a per-worker suffix, and a worker footer are all plain files you can version-control per project.
• Inspectable runs. Every run writes a losslessly replayable log, a per-worker session manifest, and each worker's stdout/stderr under a session output directory.

Who it's for

Developers who want to fan a substantial task out across several coding-agent CLIs and have one coordinator aggregate the results — while keeping control over the models, the prompts, and the audit trail.

Where to go next

• Getting Started — install the CLI and its worker prerequisites, then run your first team.
• Concepts — the coordinator/worker split, the request flow, and the tool registry.
• Configuration — config.toml, resolution order, and the prompt files.
• Workers — the structured agent-template schema and how to add your own agent type.
• Providers & Auth — OpenAI-compatible vs. Codex, and routing workers through OpenRouter.
• Reference: CLI Reference, Coordinator Tools, and Run Logs.

The source lives on GitHub and the package is published to PyPI.