An LLM Council in Your Browser – No API Keys Needed

By the AI Debate Council team · Updated October 1, 2026

One chatbot can be confidently wrong. Ask the same question to two or three different AI models and you often get two or three different answers – and no easy way to tell which one to trust. An LLM council turns that disagreement into something useful: the models answer independently, review each other, and one of them writes a final answer that says where they agree and where they don't.

This guide explains how an LLM council works, when it helps, where it doesn't, and how to run one with the AI accounts you already have.

What is an LLM council?

An LLM council is a small group of large language models that work on the same question in three stages:

1. First opinions. Each model answers the question on its own, without seeing the others.
2. Peer review. Each model reads the other answers – with the model names hidden – and points out mistakes, gaps and better arguments.
3. Synthesis. One model, the "chairman", reads everything and writes a single final answer.

The idea became widely known in late 2025 through LLM Council, an open-source project by Andrej Karpathy. In his version, each model ranks the others' answers for accuracy and insight, and a chairman model compiles the final response. Karpathy describes it as a fun weekend hack that he doesn't plan to support, and it needs an OpenRouter API key, so every question is paid per token.

Why several models beat one

• Different models make different mistakes. ChatGPT, Gemini and DeepSeek are trained by different teams on different data. An error one model makes confidently is often caught by another.
• Critique finds errors that answering misses. Research on multi-agent debate, for example Du et al., 2023, found that letting models read and respond to each other's answers over several rounds improved factual accuracy and reasoning compared with a single model.
• Hiding names reduces favoritism. When reviewers see "Expert A, B, C" instead of brand names, they judge the argument, not the source.
• Disagreement is information. A good synthesis doesn't hide the points the models still disagree on – those are exactly the points you should check yourself.

Where it doesn't help

A council is not a truth machine. Keep these limits in mind:

• Shared blind spots. If all models learned the same wrong fact, they will agree on it.
• Time. A council takes minutes, not seconds. For a quick lookup, one chatbot is fine.
• The chairman can be biased toward its own first answer. Rotating the chairman helps.
• Important decisions still need a human check – especially legal, medical and financial questions.

How to run an LLM council without API keys

The term comes from Andrej Karpathy's llm-council (November 2025): a small local app that sends one question to several models through OpenRouter, has them rank each other's answers anonymously, and lets a "chairman" model compile the result. It needs an API key and a terminal. The extension below does the same job with the chat accounts you already have, no key required.

AI Debate Council is a Chrome extension that runs the council in the browser tabs you are already signed in to:

1. Sign in to ChatGPT, Gemini and DeepSeek in Chrome, as you normally do. Free plans work.
2. Open the side panel, choose what you need – decide, plan, review or learn – and type your question.
3. The extension sends the question to each AI, then shows each one the others' anonymized answers for one to three critique rounds.
4. One AI writes the synthesis: the conclusion, where the experts agree, where they still disagree, and which point came from whom. You can ask follow-ups.

One honest caveat: automating a chat website may go against some services' terms of use, and Anthropic's terms prohibit it for claude.ai, which is why Claude is off by default in a web tab. If you'd rather avoid that risk entirely, the Pro plan can run each model through its official API with your own key – the same approach the open-source project uses.

When to use a council

• Decisions with trade-offs: a job offer, a technology choice, a large purchase.
• Plans and drafts: each model drafts, the others score and fix it, and the best parts are merged.
• Fact-checking: a claim you saw online, or an answer from a single chatbot that sounds too sure.
• Reviews: a document, a CV or a piece of code, where each model finds different problems.

For simple facts with one correct answer, a single chatbot – or a search engine – is faster.