magi-plugins

                              新  世  紀
███████ ██    ██  █████  ██    ██  █████  ███████ ██      ██  █████  ██    ██
██      ██    ██ ██   ██ ███   ██ ██      ██      ██      ██ ██   ██ ███   ██
█████    ██  ██  ███████ ██ ██ ██ ██  ███ █████   ██      ██ ██   ██ ██ ██ ██
██        ████   ██   ██ ██  ████ ██   ██ ██      ██      ██ ██   ██ ██  ████
███████    ██    ██   ██ ██   ███  █████  ███████ ███████ ██  █████  ██   ███
                           エヴァンゲリオン

MAGI SYSTEM

"One mind is a hypothesis. Three minds are a verdict."

A Claude Code plugin that brings the MAGI supercomputer council from Neon Genesis Evangelion into your terminal. Three AI agents — each powered by Claude Opus — deliberate your engineering decisions in parallel, scoring 12 dimensions simultaneously. A fourth agent, MAGI Core, synthesizes their evaluations with sycophancy detection and bias calibration.

Just as the original MAGI reflected the three facets of Dr. Naoko Akagi's personality, each agent embodies a distinct cognitive framework — with built-in cognitive resistance to prevent LLM convergence.

Installation

Plugin (Recommended)

Run inside Claude Code interactive mode:

/plugin marketplace add ohta-rh/magi-system
/plugin install magi@magi-plugins

Manual Symlink

git clone https://github.com/ohta-rh/magi-system.git
ln -s "$(pwd)/magi-system/plugins/magi/skills/magi" ~/.claude/skills/magi

Requires Claude Code.

Usage

# Full deliberation
/magi Should we migrate from REST to GraphQL?

# Comparison mode (auto-detected)
/magi React Server Components vs traditional SPA architecture

# Quick triage (single-agent, Sonnet — fast and cheap)
/magi-quick Is adding Redis caching worth the complexity?

# Code review via git diff
/magi-review

# Pre-mortem failure analysis
/magi-premortem Rewrite the auth system in Rust

# With dialectic round (inter-agent rebuttals)
/magi Should we adopt microservices? --dialectic

# With adversarial challenge (devil's advocate)
/magi Migrate to Kubernetes --adversarial

The Three MAGI

              ┌──────────────────┐
              │  BALTHASAR · 2   │
              │    「Mother」      │
              └────────┬─────────┘
                      ╱ ╲
                    ╱ MAGI ╲
                  ╱  CORE   ╲
 ┌──────────────┴──┐   ┌──────┴───────────┐
 │  CASPAR  ·  3   │───│  MELCHIOR  ·  1  │
 │   「Woman」       │   │  「Scientist」    │
 └─────────────────┘   └──────────────────┘

MELCHIOR-1 — The Scientist

"Show me the benchmark, or it didn't happen."

The cold eye of technical truth. Reasons through the Scientific Method — forming falsifiable hypotheses, seeking disconfirmation, updating beliefs with Bayesian rigor. Explicitly anti-sycophantic: states findings bluntly, actively resists the tendency to soften negative judgments.

Axis	What it measures
Correctness & Rigor	Algorithm correctness, edge cases, fault tolerance, type safety
Performance & Efficiency	Computational complexity, memory, throughput, latency, scalability
Security	Threat model, attack surface, auth design, data protection
Technical Consistency	Architecture alignment, design principles, dependency fitness

BALTHASAR-2 — The Mother

"Will the on-call engineer at 3 AM understand this code?"

The guardian of long-term survival. Evaluates across three time horizons — 1 month, 6 months, 2 years. Designs for the weakest link. Cognitive resistance: must articulate the worst-case scenario at the 2-year horizon before approving.

Axis	What it measures
Maintainability & Readability	Code comprehensibility, naming, separation of concerns
Testability	Unit/integration test ease, mockability, testable design
Operability & Observability	Logging, metrics, alerting, deployment, rollback
Team Impact	Learning curve, onboarding burden, documentation, skill alignment

CASPAR-3 — The Woman

"Correct but ugly is just a different kind of wrong."

The voice of intuition and strategic vision. Reasons through Pattern Recognition and Aesthetic Judgment — forming a gestalt impression, reading the political terrain, calculating opportunity costs. Cognitive resistance: must name the best alternative NOT chosen before approving.

Axis	What it measures
Design Elegance	API beauty, abstraction quality, DX, interface intuitiveness
Innovation & Competitiveness	Paradigm adoption, differentiation, trend alignment
Feasibility	Implementation cost, timeline realism, incremental adoption
Adaptability & Extensibility	Future-proofing, extension points, pluggability

---