🧠 BigBrain

MCP server that lets Claude Code talk to Codex (OpenAI) and Gemini (Google) CLIs. Claude stays in charge — the other models provide second opinions. No API keys needed; both CLIs use their own cached OAuth.

🍹 Origin Story

This project was born on a Saturday night vibecoding session fueled by spiced rum.

The kind of evening where you look at your API bill 💸, look at your glass 🥃, and think: "there has to be a better way."

Turns out, both Codex and Gemini have CLIs with free OAuth access. So instead of paying per token like a responsible adult, BigBrain just shells out to the CLIs like a genius cheapskate. No API keys. No billing surprises on Monday morning. Just three AI models arguing with each other over stdio while you sip your drink. 🍹

Does it work? Yes. Is it elegant? Debatable. Did it save money? Absolutely. 💰 Will your future self thank your rum-fueled past self? Also debatable. 😅

Enjoy responsibly (the rum and the code). 🥂

🏗️ Architecture

You 🫵
 └─> Claude Code (chairman — decides, writes memory, synthesizes)
       └─> BigBrain MCP Server (Python, FastMCP, stdio)
             ├─> Codex CLI  (async subprocess)
             └─> Gemini CLI (async subprocess)

🔗 Shared context: Every prompt sent to Codex/Gemini is automatically prepended with the current project's CLAUDE.md and MEMORY.md. Claude is the sole writer of these files; the other models only read them.

📍 Dynamic project detection: All tools accept an optional project_path param. Falls back to BIGBRAIN_PROJECT_PATH env var, then cwd. Works with any Claude Code project.

🛠️ Tools

👑 Council — Three-Stage Process

The most thorough tool. All three models participate equally. Based on Karpathy's llm-council:

Stage 1: Individual 🗣️       Stage 2: Peer Review 🔍      Stage 3: Chairman 🏛️
┌─────────┐                  ┌─────────┐                  ┌─────────┐
│ Claude  │──> Answer A      │ Codex   │──> Reviews &     │ Claude  │──> Final
│ Codex   │──> Answer B      │ Gemini  │    ranks all     │         │   answer
│ Gemini  │──> Answer C      │ Claude  │    (anonymized)  │         │
└─────────┘                  └─────────┘                  └─────────┘
  (all three)                  (all three)                 (chairman)

1. 🗣️ Individual — Claude forms its opinion first, then Codex and Gemini answer independently in parallel. All three answers collected.
2. 🔍 Peer Review — All answers are anonymized as Model A/B/C. Codex and Gemini each review and rank them. Claude receives the same anonymized prompt and does its own review too. No model knows which answer is whose 🕵️
3. 🏛️ Chairman — Claude synthesizes all three individual answers + all three peer reviews into the final answer.

🚀 Quick Start

1. Prerequisites

Install the Codex and Gemini CLIs, then authenticate once:

npm install -g @openai/codex @google/gemini-cli
codex   # follow OAuth prompts 🔐
gemini  # follow OAuth prompts 🔐

2. Install & Register (one command each)

pip install git+https://github.com/Leonard013/BigBrain.git
claude mcp add bigbrain -- bigbrain

3. Done 🎉

Restart Claude Code, run /mcp — bigbrain should show as connected. ✅

That's it. Start asking Claude to use ask_codex, ask_gemini, request_council, etc.

conda create -n bigbrain python=3.12 -y
conda activate bigbrain
pip install git+https://github.com/Leonard013/BigBrain.git

claude mcp add bigbrain -- /path/to/envs/bigbrain/bin/bigbrain

export BIGBRAIN_CODEX_MODEL="gpt-5.4"
export BIGBRAIN_GEMINI_MODEL="gemini-3.1-pro-preview"

git clone https://github.com/Leonard013/BigBrain.git
cd BigBrain
pip install -e ".[dev]"
pytest tests/ -v