Show HN: Boardroom MCP - Multi-advisor governance engine for AI agents
1 コメント
I actually spent a lot of time doing agentic AI before any agentic AI coding tools were available. I was building teams of agents and having them work together through copy/paste when we just had chatbots. I learned things from that exercise that still hold true today. One of the most valuable I learned is this: when using your agents to do the initial big plan for your project, have the agent who most closely matches the domain of your project create the initial plan. Then serially review the plan with your remaining agents one at a time, incorporate their feedback before sending to the next agent for review. This gets you a much greater work product than if you had them analyze and provide feedback in parallel (results in AI slop).
I noticed a recurring issue when building autonomous agents: they're great at execution but terrible at nuanced judgment. When faced with ambiguity, they just hallucinate the most statistically probable path without considering second-order effects.
Instead of trying to fix this with massive system prompts, I built an MCP (Model Context Protocol) server that offloads decisions to a multi-advisor "boardroom."
How it works: 1. Agent encounters a decision -> calls the `analyze()` MCP tool. 2. Server routes the query to relevant advisors (from 38 domains, 450+ profiles). 3. Advisors debate. This is the core mechanic: if advisors agree too quickly, the system flags it. Tension is mandatory. 4. The server synthesizes the debate, logs it to a persistent LEDGER (so the agent has institutional memory), and returns a risk-scored recommendation.
I built this locally so there's zero cloud dependency and your agent data stays yours. It works natively with Claude Desktop, Claude Code, Cursor, Windsurf, or any MCP client.
You can checkout the core engine here: https://github.com/randysalars/boardroom-mcp If you want to read the docs: https://salars.net/boardroom/docs
Would love to know if anyone else has experimented with structured multi-agent debate vs LLM-as-a-judge patterns.