A confident-sounding answer from one AI model isn't the same thing as a correct one. The usual way to catch that gap is manual: paste the same question into ChatGPT, then Claude, then Gemini, and eyeball the differences yourself. It works, but it's tedious enough that most people just don't do it — and the questions where it matters most (higher stakes, more ambiguity) are exactly the ones people skip the check on. Maestro automates that comparison instead of relying on you to remember to do it.
The problem
Every model has blind spots, and they're not the same blind spots. Two models agreeing on a fact is a much stronger signal than one model stating it alone, but nothing in a typical chat interface surfaces that distinction — you get one answer, delivered with the same tone of confidence whether it's well-supported or a guess. Cross-checking against other models fixes this, but only if someone actually does it, every time, for every claim that matters.
How it's built
Maestro is a single HTML file — no build step, no backend, no server to maintain — that talks directly to OpenRouter from the browser, giving it access to 46+ models across 20 providers. It has four modes:
- Auto — a cheap classifier model reads the question's complexity and stakes, then picks how many models to involve.
- Solo — one model, for quick, low-stakes questions where cross-checking is overkill.
- Duel — two models answer independently, useful for a fast sanity check.
- Council — three to five models weigh in, for the questions where getting it wrong actually costs something.
Whichever mode is used, the models that answer do so independently — they don't see each other's responses. A separate judge model then compares the independent answers, synthesizes a consensus response, and explicitly flags any claim that only one model made as a possible hallucination rather than quietly blending it in as fact.
Everything — the API key, configuration, and history — is kept in the browser's localStorage.
There's a live model-availability checker, a provider-tabbed model picker, and an analytics dashboard
that tracks spend and savings versus always defaulting to the most expensive frontier model. The
whole thing is deployed statically on GitHub Pages.
Why the architecture matters
The zero-backend design isn't just a convenience — it means your API key and your questions go straight from your browser to OpenRouter and never pass through a server you don't control. It also means the routing logic (which models to use, when) is the only place where cost discipline lives: cheap questions stay cheap, and the expensive multi-model cross-checking only kicks in when the classifier decides the stakes justify it.
Try it
Maestro is live at soroush-thr.github.io/Maestro-AI, and the code is on GitHub. If you've ever caught yourself trusting an AI's answer a little too quickly, it's worth a look.