fshiori

fshiori / magi

Public

Three LLMs debate to make better decisions than any single model. Inspired by EVA's MAGI supercomputer. Critique mode: 88% vs single Claude Sonnet's 76%.

36
4
100% credibility
Found Mar 28, 2026 at 36 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

MAGI is a tool that queries three AI language models from different perspectives, letting them vote, debate, or critique to produce a single high-quality decision with confidence scores and dissenting views.

How It Works

1
📖 Discover MAGI

You hear about MAGI, a clever system where three AI thinkers with different views debate your questions for smarter answers.

2
🛠️ Set it up

You add MAGI to your computer in moments, like installing a helpful app.

3
🔗 Link the thinkers

You connect three AI services so they can share ideas and work as a team.

4
Ask a question

You type in your question, like 'Should we try this new idea?', and watch the three AIs vote, debate, or critique each other.

5
Choose your style
Quick vote

Get a fast majority decision when everyone agrees easily.

🗣️
Full debate

See them challenge each other's ideas round by round for higher quality.

🤖
Smart auto

MAGI picks the right approach based on how much they disagree.

6
📊 View the dashboard

Open a cool visual screen to watch the AIs think live, like a command center.

🎉 Wise decision delivered

You receive the final answer with confidence level, minority opinions, and full story of how they agreed or changed minds.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 36 to 36 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is magi?

Magi pits three LLMs in structured debate—vote for quick majority, critique for multi-round revisions, adaptive for auto-escalation—to output a decision dossier with ruling, confidence score, minority report, and full trace. Python-based via LiteLLM for 100+ providers like OpenRouter, it beats single-model limits: cheap big three llms in critique mode score 88% on benchmarks vs. Claude Sonnet's 76%. CLI commands like `magi ask "query"`, `magi diff --staged` for code review, `magi judge`, plus NERV web dashboard for live visualization.

Why is it gaining traction?

Unlike github three way merge or simple voting in github three body problem setups, Magi tracks mind changes, generates dissent analysis, and uses presets (code-review, strategy) for targeted debates. Fault-tolerant: one model down? Degrades gracefully. Hook: `magi diff` delivers security/perf/quality critiques from divergent personas, with traces for `magi replay`—practical over raw ensembles.

Who should use this?

DevOps engineers reviewing PR diffs for bugs/security. PMs querying "monolith vs microservices?" with strategy preset. AI hobbyists benchmarking magi system github vs. magi ai github alternatives like github three js game decision sims. Ideal for CLI-driven experimentation, not high-volume APIs.

Verdict

Worth pip-installing for `magi diff` and dashboard fun—docs shine, 83 tests cover edges, MIT license. But 36 stars and 1.0% credibility signal early days; prototype quality, not production-ready. Try if magi v2 github intrigues.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.