bolivartech-plugins

MAGI — Multi-Perspective Analysis Plugin for Claude Code

A Claude Code plugin that implements a multi-perspective analysis system inspired by the MAGI supercomputers from Neon Genesis Evangelion.

Three specialized AI agents independently analyze the same problem from complementary — and deliberately adversarial — perspectives, then synthesize their verdicts via weight-based majority vote.

---

Why Three Adversarial Perspectives?

The MAGI in Evangelion

In Neon Genesis Evangelion (1995, Hideaki Anno / Gainax), the MAGI are three supercomputers that govern Tokyo-3's critical decisions. Each embodies a different facet of their creator, Dr. Naoko Akagi: Melchior (the scientist), Balthasar (the mother), and Caspar (the woman). Decisions require consensus — no single perspective dominates.

This design reflects a profound insight: complex decisions benefit from structured disagreement. A single decision-maker, no matter how capable, carries blind spots. Three independent evaluators with different priorities surface risks, trade-offs, and opportunities that any one of them would miss.

The Theory in Practice

The adversarial multi-perspective model addresses well-documented cognitive biases in software engineering:

Bias	How MAGI Mitigates It
Confirmation bias	Three agents with different evaluation criteria are unlikely to share the same blind spots
Anchoring	Agents analyze independently — no agent sees the others' output before forming its own verdict
Groupthink	Caspar (Critic) is designed to be adversarial; its role is to find fault, not agree
Optimism bias	The weight-based scoring penalizes reject (-1) more heavily than approve (+1), making negative signals harder to override
Status quo bias	Each agent evaluates from first principles against its own criteria, not against "how things are done"

The key insight is that disagreement between agents is a feature, not a failure. When Melchior (Scientist) approves but Caspar (Critic) rejects, the dissent surfaces a genuine tension between technical correctness and risk tolerance. Unanimous agreement on non-trivial input may indicate insufficiently differentiated prompts, not actual consensus.

In practice, the system works best for decisions with:

• Genuine uncertainty — multiple valid approaches exist
• Significant consequences — the cost of a wrong decision is high
• Hidden trade-offs — benefits and risks are not immediately obvious

For trivial questions with one clear answer, the complexity gate skips the full system and responds directly.

---

Documentation

For the full technical reference, see docs/MAGI-System-Documentation.md.

---

Agents

Agent	Codename	Lens	Personality
Melchior	Scientist	Technical rigor and correctness	Precise, evidence-based, favors proven solutions
Balthasar	Pragmatist	Practicality and maintainability	Grounded, trade-off oriented, advocates for the team
Caspar	Critic	Risk, edge cases, and failure modes	Adversarial by design, finds what others miss

---

Installation

From GitHub (for users)

# 1. Add this repo as a marketplace source
/plugin marketplace add BolivarTech/magi

# 2. Install the plugin
/plugin install magi@bolivartech-plugins

# 3. Use it
/magi

To update after new versions are published:

/plugin marketplace update

Local Development

# Option 1: Plugin flag
claude --plugin-dir /path/to/magi

# Option 2: Symlink for auto-discovery (no flags needed)
mkdir -p .claude/skills
ln -s ../../skills/magi .claude/skills/magi
claude

Changes are picked up with /reload-plugins without restarting.

---

Usage

Invoke with /magi or natural trigger phrases:

MAGI review this code
Give me three perspectives on this design
MAGI analysis of this problem

Modes

Mode	When to Use	Example
code-review	Reviewing code or diffs	"MAGI review this PR"
design	Evaluating architecture decisions	"MAGI analyze this migration plan"
analysis	General problem analysis, trade-offs	"MAGI should we use Redis or Postgres for this?"

CLI (Direct Execution)

python skills/magi/scripts/run_magi.py <mode> <file_or_text> [--model opus] [--timeout 300] [--output-dir <dir>]