MAGI

A CLI that puts a decision to a vote across three LLMs — Claude, Gemini, and ChatGPT — styled after the MAGI supercomputer system from Neon Genesis Evangelion. Each model is one independent unit (Melchior-1 / Balthasar-2 / Casper-3); a verdict requires a 2-of-3 majority, or the system reports a deadlock.

Install

git clone <your-repo-url>
cd magi
python -m venv .venv && source .venv/bin/activate
pip install -e .

Or, once published to PyPI:

pipx install magi-cli

Configure

Copy .env.example to .env and add your API keys:

cp .env.example .env

ANTHROPIC_API_KEY=sk-ant-...
GEMINI_API_KEY=...
OPENAI_API_KEY=sk-...

Usage

magi decide "Should we migrate the pipeline to Rust?"

magi decide "Rust or Go for the new service?" --options Rust,Go

magi units          # show which model is wired to which unit

Running decide boots the system (an ASCII "MAGI" banner + a per-unit ONLINE handshake), then shows a live hexagonal diagram — Balthasar on top, Casper and Melchior below, all converging on a central MAGI hub — with each node pulsing and spinning while its model is actually thinking. Once resolved, it reveals each unit's full vote with reasoning, then a final verdict panel — CONDITION GREEN for a clear 2/3 majority, CONDITION RED for a deadlock. The hex diagram needs at least ~70 terminal columns; narrower terminals automatically fall back to a plain vertical list.

Pass --plain (or set NO_COLOR=1 / MAGI_PLAIN=1) to strip all of that down to plain text — for scripts, CI, or anyone who just wants the answer.

Notes

• Model IDs move fast. The defaults in src/magi/config.py (claude-sonnet-5, gemini-3.1-pro-preview, gpt-5.6-sol) were current as of August 2026. If a call fails with a "model not found" error, check each provider's current model list and override via MAGI_CLAUDE_MODEL / MAGI_GEMINI_MODEL / MAGI_OPENAI_MODEL env vars, or edit config.py directly.
• Reassign units by editing the UNITS tuple in config.py — swap which provider sits behind Melchior/Balthasar/Casper, or add a fourth/fifth unit (the tally logic generalizes; the "2-of-3" framing in render_verdict is the only MAGI-specific part).
• Publishing to PyPI: pip install build twine && python -m build && twine upload dist/* — see the project's earlier chat thread for the full walkthrough.

License

MIT