Nyambot is the MCP connector that gives Claude, ChatGPT, Le Chat, Copilot and Gemini direct access to official French public databases: real estate prices actually recorded at the land registry (DVF), the company registry, the legal code, health providers, job listings, benefits and administrative procedures. Not estimates — the date, the amount, the source. Every figure is re-verifiable. Connect in 60 seconds, or self-host from PyPI.
Hi Product Hunt 👋
I'm Gabriel, solo builder from France.
What started this. I kept watching people ask ChatGPT what an apartment was worth in a French town, and get a confident, invented number. Not because the data doesn't exist — France publishes every property sale recorded at the notary, the full company registry, the entire legal code. It's all open. It's just unusable without a week of plumbing per source.
The problem. Open doesn't mean usable. Every source has its own API, its own format, its own identifiers, its own traps. Wiring three of them into an agent is a week of work, and maintaining them is forever. So almost nobody does it, and assistants keep guessing about France.
What Nyambot is. One MCP connector, 21 tools. You paste a URL into Claude, ChatGPT, Cursor, Le Chat or Copilot, and your assistant answers with the actual figure, dated, with a link back to the official source.
How my thinking changed. I started out building retrieval — find the right document, hand it to the model. That was wrong. The questions people actually ask can't be retrieved. "What's this 24 m² flat worth?" has no answer written anywhere. You have to pull every sale signed within 200 metres, filter by comparable surface and period, and compute a median. The answer doesn't exist until someone asks it. That moved the whole project from search to computation.
The moment that taught me the most. Last week my own tool flagged a bug in itself, in the middle of an answer. French land registry files list one row per lot, not per sale, and the total price is repeated on every lot. Without grouping, a median for studios came out at €6,251/m² instead of €2,330 — and filtering by surface made it worse, not better, because it selects exactly the small lots of bulk building sales. That bias is quietly sitting in a lot of tools built on this dataset. I found it, measured it (10.6% of sales aren't attributable), fixed it, shipped it.
That's the actual work. Not accessing open data — understanding it well enough to be trusted with it.
Free to connect, no credit card. Self-host from PyPI if you'd rather keep it in-house.
Tell me what breaks. I'm answering all day.