🚀 I built an open-source tool that lets one AI coding agent 🧠 plan while another writes the code.
I named it Multiharness. Using this, the planner stays in control of review, and the implementer does the typing.
How it works:
• It runs as an MCP server 🔌 inside the planning agent, with tools to delegate work, wait on it, and cancel it
• The implementer runs headless as a background job with tracked status, session and logs
• Every job comes back with independent Git evidence, so the planner reviews the actual repository instead of trusting the agent's own summary
• Fixes go back to the same implementer session, so repair rounds keep their context
• Timeouts and cancellation stop the whole process tree
Supported agents: Claude Code, Codex CLI, OpenCode and Command Code as planner or implementer, and Muse Code as implementer.
That allows setups like:
• Claude Code plans, Codex CLI implements
• Claude Code plans, Command Code implements
• Claude Code plans on one model and implements on Sonnet, on a single subscription
There's also an optional routing step: a small decision model (Jev or Laya) decides whether a request needs a plan at all, so trivial changes skip planning. ⚡
It installs with one line on Windows, macOS and Linux, and a setup wizard handles the rest: installing missing agents, opening sign-ins, and connecting itself to your planner.
Each agent is a single adapter file, and the test suite drives every agent, the MCP server and the wizard against fake CLIs. 🧪
Repo: https://github.com/ErOr-0/Multiharness-Core
Website: https://multiharness.mdfahimhossen.space
⭐ Give it a star if you find it useful!
Top comments (0)