Multi-model AI collaboration for Claude Code. Get consensus-driven code reviews and second opinions from multiple AI models in parallel.

Commands

Command	Purpose
/review	Multi-model code review on your staged changes
/consult	Get a second opinion when stuck on a problem

How It Works

/review                              /consult "why is this broken?"
   │                                    │
   ├── Codex ───► reviews diff          ├── Codex ───► reads conversation
   ├── Gemini ──► reviews diff          ├── Gemini ──► reads conversation
   ├── Claude ──► reviews diff          ├── Claude ──► reads conversation
   └── ...                              └── ...
   │                                    │
   ▼                                    ▼
   Consensus: issues flagged            Synthesis: consensus suggestions,
   by 2+ models highlighted             unique perspectives, disagreements

Why multiple models?

• Different training data → different blind spots
• 2+ models flagging the same issue → stronger signal
• Diverse perspectives surface better solutions

Inspiration

Inspired by LLM Council — the idea that multiple LLMs reviewing the same problem surfaces stronger signals than any single model.

Installation

git clone https://github.com/caiopizzol/conclave ~/dev/conclave
cd ~/dev/conclave
bun run register

To unregister:

bun run unregister

Configuration

Tools (~/.config/conclave/tools.json)

{
  "persistence": {
    "enabled": true,
    "required": false,
    "data_dir": "~/.local/share/conclave/reviews"
  },
  "tools": {
    "codex": {
      "enabled": true,
      "scope": ["review", "consult"],
      "command": "codex exec --full-auto -",
      "model": "gpt-5.2-codex",
      "description": "OpenAI Codex CLI"
    },
    "claude-opus": {
      "enabled": true,
      "scope": ["consult"],
      "command": "claude --print",
      "model": "opus",
      "description": "Claude Code (Opus)"
    },
    "gemini": {
      "enabled": true,
      "scope": ["review", "consult"],
      "command": "gemini -o text",
      "model": "gemini-3-pro-preview",
      "description": "Google Gemini CLI"
    }
  },
  "prompts": {
    "review": "~/.config/conclave/prompt.md",
    "consult": "~/.config/conclave/consult-prompt.md"
  }
}

Tool Fields

Field	Required	Description
enabled	Yes	Whether to use this tool
scope	No	Array of commands: ["review"], ["consult"], or ["review", "consult"]. If omitted, tool is used for all commands
command	Yes	CLI command to run
model	No	Model to use (injected via --model or -m flag)
description	No	Human-readable description

You can define multiple entries for the same provider with different models (e.g., claude-opus and claude-sonnet).

Persistence Fields

Field	Default	Description
enabled	true	Save review results to disk for later analysis
required	false	If true, halt on persistence failure instead of continuing
data_dir	~/.local/share/conclave/reviews	Directory for saved review data

Review results are saved as JSON files containing raw model outputs, timestamps, and investigation results. This enables:

• Tracking review history across PRs
• Analyzing model performance over time
• Recovering from context compaction

Supported models:

Tool	Models	Documentation
Codex	gpt-5.2-codex, gpt-5.1-codex-mini, gpt-5.1-codex-max, gpt-5.2	Codex Models
Claude	opus, sonnet, haiku (aliases) or full names like claude-opus-4-5-20251101	CLI Reference
Gemini	gemini-2.5-pro, gemini-2.5-flash, gemini-3-pro-preview, gemini-3-flash-preview	Gemini CLI
Qwen	coder-model (default), vision-model	Qwen Code Docs
Mistral	Config-based (~/.vibe/config.toml)	Mistral Vibe Docs
Grok	grok-code-fast-1, grok-4-1-fast-*, grok-4-fast-*, grok-3, grok-3-mini	xAI API Models
Ollama	qwen3-coder:480b-cloud, devstral-2:123b-cloud, or any model from library	Ollama Library

Note: Ollama cloud models use :cloud suffix and require OLLAMA_API_KEY environment variable. Get your API key at ollama.com. You can also run local models (e.g., qwen2.5-coder:7b), but they are slow and require significant memory (~8GB+ RAM for 7B models).

Note: Mistral and Grok use command-line argument passing (not stdin), which has a ~200KB limit on macOS. Very large diffs may cause these tools to fail while other tools succeed.

Note: Grok uses the community CLI (@vibe-kit/grok-cli) until xAI releases the official "Grok Build" CLI.

Prompts

Customize prompts for each command:

File	Template Variables
~/.config/conclave/prompt.md	{{branch}}, {{target_branch}}, {{diff}}
~/.config/conclave/consult-prompt.md	{{history_file}}, {{question}}, {{cwd}}

Authentication

Tool	Install
Codex	npm install -g @openai/codex
Claude	Built-in
Gemini	npm install -g @google/gemini-cli
Qwen	npm install -g @qwen-code/qwen-code
Mistral	pipx install mistral-vibe
Grok	bun add -g @vibe-kit/grok-cli; export GROK_API_KEY="key" in ~/.zshrc
Ollama	ollama.com/download; cloud: export OLLAMA_API_KEY="key" in ~/.zshrc; local: ollama pull <model>

Usage

Code Review

# Review staged changes
git add -p
/review

Second Opinion

# When stuck on a problem
/consult "why is the table rendering broken after paste?"

# When going in circles
/consult "I've tried X, Y, Z but none work"

# Validate an approach
/consult "is this the right way to handle state here?"

/consult passes your Claude Code conversation history to external models, so they can see what's already been tried and avoid suggesting the same things.

Philosophy

More models ≠ better. The value is consensus:

• 1 model flags issue → might be noise
• 2+ models flag same issue → likely real
• Different perspectives → surface blind spots

Conclave surfaces what matters.

Workflow

/review follows a state machine pattern with checkpoints to ensure reliability:

[INIT] → [GATHERING] → [SPAWNING] → [TOOLS_COMPLETE]
                                          │
                                    ⛔ CHECKPOINT
                                          │
                                    [PERSISTED] → [SYNTHESIZING] → [INVESTIGATING] → [COMPLETE]

The ⛔ checkpoint ensures review data is persisted before synthesis. If persistence.required is true, the workflow halts on persistence failure.

License

MIT