Fusion Council · your models, our conductor

Ask every AI at once.
Get one verdict.

One hard question goes to a panel of models — your own Claude Code and Codex exploring the repo read-only, plus any catalog model through your OpenRouter key. A chairman judges the answers against a rubric and folds them into one verdict. You're pinged only when they disagree.

Download Backgrind How the pattern works →

Three steps to a judged answer.

1

Ask something hard

An architecture call, a migration plan, a review of this change — one question every seat answers independently, blind to each other.

2

Pick the panel

2–4 seats, any combination: your CLIs (enforced read-only in the repo) and any OpenRouter model from a searchable catalog. Presets pick sensible line-ups.

3

The chairman judges

Answers are anonymized and scored against a rubric — no brand bias. The verdict lands with AGREEMENT and CONFIDENCE ratings, and one click sends it to your active agent.

Two kinds of seats. Zero new trust.

CLI seats — deep, in your repo

Claude Code (--permission-mode plan) and Codex (--sandbox read-only) explore the codebase with their write paths flag-disabled. Same login, same subscription you already pay for. CLIs without a verified read-only flag don't get a seat — fail closed.

API seats — any model, your key

Paste your OpenRouter key and the whole catalog can sit on the panel — GPT, Gemini, DeepSeek, Llama. Calls go straight from your machine to openrouter.ai; nothing touches our servers, and the panel says honestly that API seats answer without repo access.

Go deeper.

Questions people ask.

What is a model council?
Several AI models answer the same question independently, 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 vendors break the correlated blind spots a single model (or the same model re-rolled) cannot escape.
Which models can sit on the panel?
Two kinds of seats. CLI seats run the agents you already have — Claude Code and Codex CLI, each with an enforced read-only mode so they can explore your repo but never change it. API seats use your own OpenRouter key: any catalog model — GPT, Gemini, DeepSeek, Llama and hundreds more — can sit on the panel or chair it.
Can the council see my code?
CLI seats explore the repo in enforced read-only mode on your machine. API seats get no repo access at all — they reason from your question alone, and the panel says so honestly. API calls go straight from your machine to openrouter.ai; nothing passes through our servers, and your key never leaves the device.
What does it cost?
Nothing extra to us. CLI seats bill the subscriptions you already pay for (Claude, ChatGPT). API seats bill your own OpenRouter key at provider prices — a council round typically costs cents. Backgrind sells the conductor, not the tokens.
When should I convene a council instead of just asking my agent?
When being wrong is expensive: architecture calls, migration plans, security-sensitive changes, reviews of large diffs. For trivial edits a single agent is fine — the council is the second opinion for decisions you would otherwise sleep on.
What happens when the models disagree?
Disagreement is the signal. The verdict carries an AGREEMENT (consensus / split) and CONFIDENCE (high / low) rating; Backgrind pings you only on a split or low confidence. Consensus lands as a quiet toast — no interruption.
Convene your first council