llm-council, in plain language

What is this?

Imagine you have an important question and four smart friends. Instead of asking just one friend and hoping they're right, you:

1. Ask all four separately, so nobody copies anybody.
2. Show each friend everyone's answers with the names hidden and ask them to grade the answers and rank them best to worst.
3. Give everything — answers, grades, and who wrote what — to a chairman, who writes the final, best answer and explains where people disagreed.

That's exactly what this tool does, except the "friends" are AI models: Claude, ChatGPT/Codex, Gemini, and a model that runs entirely on this PC.

Why bother with a whole council?

Any single AI can be confidently wrong. When four different AIs from four different companies agree, you can trust the answer a lot more. When they disagree, the chairman has to say who's right and why — which surfaces the disagreement instead of hiding it.

What do I get?

• The chairman's final answer, printed right in the terminal.
• A complete transcript (every answer, every grade, the verdict) saved as a file in the reports/ folder, so you can always check the council's work.

What does it cost?

No extra API bills. Each AI is used through its normal app subscription that's already installed and signed in on this computer. There are no passwords or secret keys stored in this project at all.

What if one of the AIs is down?

The meeting goes on without them. If the chairman itself fails, the tool keeps everyone's answers and grades anyway and tells you to pick the final answer yourself — nothing is lost.

How do I use it?

Usually you don't run it by hand — Claude Code does, via the /llm-council command ("ask the council") or /council-code-review (the council reviews code changes before they're committed). But by hand it's one line:

llm-council "Should I use SQLite or Postgres for a desktop app?"