BackgrindYou ask once
20–500 characters of plain language — no code dumps. The demo panel has no repo access, so it reasons from your question alone and the verdict says so honestly.
Free browser demo · no signup
Ask three AI models. Get one verdict.
One hard engineering question, answered by one model each from OpenAI, Anthropic and MiniMax, blind to each other. A Google Gemini chairman then judges those answers into a single verdict, rated for agreement and confidence. About 45 seconds, no account, no API key.
- Three vendors, three independent answers
- Rated consensus / split · high / low
- Shareable permalink for the verdict
Your question
3 models · 1 chairman · 1 verdict
On the panel OpenAIAnthropicMiniMax Chairman · Google Gemini
One model each — their fast tiers, on our dime. The verdict page names the exact model behind every answer.
0 / 500 characters (min 20)
Council in session 0:00
- Panel — three models answering independently ~30s
- Synthesis — the chairman weighs the answers ~15s
- Verdict — agreement + confidence
Keep this tab open — the verdict opens on its own permalink.
Three questions, three verdicts — that is the whole free demo.
You have seen what a council does. Backgrind convenes the same kind of panel on the agents you already pay for — Claude Code, Codex, Gemini — reading your real repository, read-only, with no run limit.
Not sure what to ask? Take one of these.
Click a card and the question drops into the box — edit it or send it as is.
Free demo limits, stated plainly: 3 questions per visitor and a shared daily budget · each vendor's fast tier, not its flagship · 20–500 characters of plain language · your verdict gets a public link that anyone can open, kept for a year — a month if you unlist it, and the page shows the date.
Three answers, then a judgement.
Not an ensemble average and not a merge of three texts. The panel answers first, the chairman rules second — and it has to say out loud how much the panel agreed.
Three models answer blind
One model each from OpenAI, Anthropic and MiniMax runs in parallel — their fast tiers, on our dime. None of them sees another answer, so their blind spots do not line up the way one model re-rolled twice always does.
A chairman judges, not merges
The answers arrive anonymized as Expert A, B and C — no brand wins on reputation. The chairman must open with AGREEMENT (consensus or split) and CONFIDENCE (high or low), then rule.
What lands when the 45 seconds are up.
Every run returns the same artifact: two ratings, one verdict, and the three raw answers underneath — on a link you can send to whoever disagreed with you.
Sample layout
Should a new SaaS backend in 2026 start as a modular monolith or as microservices?
Agreement consensus Confidence high
The chairman's ruling goes here — one paragraph that commits to an answer, then the key points and the trade-off the panel considered decisive.
The panel — three independent answers, kept underneath
backgrind.com/council/r/… · public for a year, then deleted
An illustration of the output format, not a saved result — every verdict is written by the panel that answered your question.
Three models that never saw each other's work landed in the same place. That is the closest thing to a second opinion you can get in 45 seconds.
The disagreement is the finding. It marks the exact decision that deserves a human — which is the only moment the desktop app interrupts you.
What the demo can't do — and the desktop app can.
This page is deliberately small: each vendor's fast tier, three questions per visitor, and no idea what your code looks like. The product it advertises has none of those constraints.
This browser demo
Free, honest, and limited on purpose.
- One model each from OpenAI, Anthropic and MiniMax — their fast tiers, on our dime
- A Google Gemini chairman, and every answer printed with its exact model id
- 3 questions per visitor, plus a daily budget for everyone
- No repo access: the panel answers from your question alone
- The verdict lives on a public link, kept for a year — a month if you unlist it
- No account, no install, nothing to configure
Backgrind desktop — Fusion Council
Your models, your repo, your machine.
- Convenes the frontier agents you already pay for — Claude Code, Codex, Gemini
- They explore your actual repository in enforced read-only mode
- Any OpenRouter model can take a seat, on your own key
- No question limits — your subscriptions pay the tokens, we sell the conductor
- Your code and your keys never touch our servers
- Pings you only on a split or low confidence; consensus lands as a quiet toast
Council demo FAQ
What is a model council?
Several AI models answer the same question independently, blind to each other, then a chairman model judges the answers against a rubric and folds them into one verdict with an agreement and confidence rating. Independent answers from different models break the correlated blind spots a single model cannot escape.
Which models sit on the demo panel?
One model each from OpenAI, Anthropic and MiniMax, with a Google Gemini chairman judging their answers. Four vendors, on our dime and without a key of yours. We run each vendor’s fast, inexpensive tier rather than its flagship — and you do not have to take our word for it: the result page prints the exact model id next to every answer.
Is this demo really free?
Yes — a model each from OpenAI, Anthropic and MiniMax on our dime, no account, no signup. It is limited to 3 questions per visitor and a daily budget for everyone. The Backgrind desktop app has no such limits: it convenes the frontier models your own subscriptions already pay for.
Why only 3 questions?
Because every run spends real money on our side, against a fixed daily budget. Three questions is enough to see whether a council changes your mind, and it keeps the demo open to everyone instead of empty by lunchtime. The desktop app has no limit because the panel runs on the subscriptions you already pay for.
How is the desktop Fusion Council different?
The demo models answer from your question alone. The desktop app convenes your own CLIs — Claude Code, Codex, Gemini — exploring your actual repository in enforced read-only mode, so the panel reasons about your real code. Your key and your code never touch our servers.
Can I share a result?
Every verdict gets a permalink you can send to a teammate, and anyone holding that link can open it. Clear the "Let Backgrind share this result" box before you ask and the run stays unlisted instead: the link still works, but we never surface it and it no longer previews when pasted into a chat. A shared result is kept for a year, an unlisted one for a month, and the result page prints its own deletion date.
What happens to my question?
The question and the model answers are stored to serve the shareable result page — for a year if you leave the share box ticked, for a month if you unlist the run — then deleted. No account or personal information is collected. Details are in the privacy policy.
Go deeper: the Fusion Council, how to ask multiple AI models one question and what model consensus actually tells you.
In the desktop app
Stop shipping on one model's confidence.
Backgrind convenes the agents you already pay for on the decisions that are expensive to reverse — reading your real repository, read-only, and pinging you only when they disagree. Ask this demo first; install it when the verdict earns it.