Agent Council

Agent Council

한국어 버전 (Korean)
A skill that gathers opinions from multiple AI CLIs (Codex, Gemini, ...) and lets a configurable Chairman synthesize a conclusion.
Inspired by Karpathy's LLM Council

Key Difference from LLM Council

No additional API costs!

Unlike Karpathy's LLM Council which directly calls each LLM's API (incurring costs), Agent Council uses your installed AI CLIs (Claude Code, Codex CLI, Gemini CLI, ...). This is especially useful if you mainly use one host CLI and occasionally consult others via subscriptions.

Skills are much simpler and more reproducible than MCP. We recommend installing via npx and customizing it yourself!

Demo

https://github.com/user-attachments/assets/c550c473-00d2-4def-b7ba-654cc7643e9b

How it Works

Agent Council implements a 3-stage process for gathering AI consensus:

Stage 1: Initial Opinions
All configured AI agents receive your question simultaneously and respond independently.

Stage 2: Response Collection
Responses from each agent are collected and displayed to you in a formatted view.

Stage 3: Chairman Synthesis
Your host agent (Claude Code / Codex CLI / etc.) acts as the Chairman by default (role: auto), synthesizing all opinions into a final recommendation. Optionally, you can configure a Chairman CLI command to run synthesis inside council.sh.

Setup

Option A: Install via npx (Recommended)

This copies the skill files to your current project directory.
If you upgrade Agent Council and hit a runtime error like Missing runtime dependency: yaml, re-run the installer command above to refresh your installed skill files.

By default, the installer auto-detects whether to install for Claude Code (.claude/) and/or Codex CLI (.codex/) based on what’s available on your machine and in the repo.

Installed paths:
-…