Hello HN, Inspired by Andrej Karpathy's "LLM Council" concept, I built a lightweight MCP (Model Context Protocol) server that lets different AI agents communicate with each other locally. Most multi-agent frameworks require complex orchestration or heavy installs. I wanted something simple that feels like it "just works". The Problem: I often find myself manually copy-pasting context between coding agents (like Claude Code to Codex or vice versa) to get a second opinion. The Solution: Agents Council is an MCP server that acts as a shared message bus. It runs locally via stdio and allows any MCP-compliant agent to "summon" other agents into a council session to debate solutions or review code. Key Features: Zero Install: Runs directly via npx agents-council@latest. Summon other agents: agents can also summon other agents directly (for the moment only Claude Code is supported) Privacy First: Works 100% offline if you are using local models; state is stored in a local JSON file. Protocol Agnostic: Works with any client that supports MCP (Gemini, Cursor, etc.). It’s open source and I’d love your feedback on the implementation or ideas for new integrations. |