MAGI

MAGI SYSTEMS — DECISION SUPPORT

質問 / QUESTION · 審議 / DELIBERATION · 解決 / SOLUTION

A small Evangelion-inspired multi-agent decision system powered by local LLMs via LM Studio.

The MAGI

MELCHIOR-01 — The Scientist: evidence, logic, truth, causality.

BALTHASAR-02 — The Mother: consequences, risk, responsibility, practicality.

CASPER-03 — The Woman: human behavior, motivation, perception, social context.

How it works

All three independently evaluate the proposition.
If they disagree, they see the previous round's judgments.
Each MAGI may defend or change its position.
This repeats for MAX_ROUNDS.

Unanimous agreement produces CONSENSUS.
If the final round still has disagreement, the 2–1 majority becomes the final decision.

The LLMs provide the reasoning and votes; the Python engine controls the deliberation protocol and final vote.

MAGI DECISION SUPPORT SYSTEM

質問 — PROPOSITION / QUESTION

Status

🟢 Working demo

• Local LM Studio models
• 3-agent MAGI deliberation
• Multi-round decision process
• Consensus + majority determination
• Individual reasoning + confidence
• Evangelion MAGI-inspired CRT UI
• Live deliberation status
• Configurable rounds and response limits
• Responsive interface

Setup venv

First, setup your .env file (theres a .env.example file in the repo)

MODEL_ID=your-model-id
OPEN_AI_BASE_URL=http://localhost:1234/v1
OPEN_AI_API_KEY=your-api-key

MAX_ROUNDS=5
MAX_RESPONSE_TOKENS=300

Python 3.14 needed

# Initial setup only
uv venv
source .venv/bin/activate
uv sync

Run

Add your config to the .env file first!

Start the LM Studio local server.

Then start the APP server:

uvicorn app:app --reload --port 8765

## If not in active venv: 
# uv run uvicorn app:app --reload --port 8765 

Web interface will be at http://localhost:8765/

TODO

• More advanced consensus logic
• Even more MAGI aesthetics
• Different API for each MAGI node