<div align="center">

<img src="docs/screenshots/cli-session.gif" alt="CodeJury in the terminal: which model owns each stage, an indexed repository, a request as tickets, the Planner's verified plan, and a four-judge jury with each juror's verdict" width="900">

_{One unbroken session against <a href="https://github.com/go-gitea/gitea">go-gitea/gitea</a> — Go, TypeScript and templates, 120,521 symbols indexed. Which model owns each of the six stages, the request that started it, the ticket it became, the Planner's plan pinned to real symbols, then the jury: four judges on four different providers, each with its own verdict, one of them dissenting. Every screen is real state from a delivery that actually ran; nothing here is a mockup.}

<br>

### A terminal-first coding agent that is reviewed by a **jury**, not by a judge.

**One LLM grading another LLM's code is not a review — it is a coin flip with a confident voice.** So CodeJury sends every change to a panel of independent, differently-modelled judges and a foreperson who synthesizes one verdict. And because a panel costs tokens, the agent earns them back: it navigates your repository through a **persistent code graph plus semantic search over graph nodes**, so it *looks up* where a change goes instead of burning context rediscovering it.

**And nothing about it is fixed.** Six stages — *knowledge, PM, planner, dev, QA, review* — and you choose the provider and model for every one of them, live, from the terminal. Claude plans, Codex writes, GPT reviews. The jury itself is a roster you seat, re-model and re-brief yourself. All of it runs on the coding CLIs you're already logged into, so **you may not need an API key at all**.

<br>

`pip install` · one command to run · no API key required · **macOS · Linux · Windows**

[](https://github.com/krishagarwal314/CodeJury/actions/workflows/ci.yml) ![Python](https://img.shields.i