Frequently asked questions

Curator turns plain-English questions about your data into real Python, runs it on your actual rows, and shows you both the answer and the exact code that produced it. It’s built for the part of analysis that matters: trusting the result.

You drop in a dataset and ask a question. Curator writes the Python to answer it, executes that code in a sandbox right in your browser (via Pyodide/WebAssembly), and returns the answer with a chart. Open “Show your work” for the exact code and the reasoning behind it, or switch to the Table and Raw views. Don’t like a line? Edit it and re-run.

No. It runs entirely in your web browser — nothing to download or set up.

Yes — every answer carries its Python one click away under “Show your work.” You can change a line (different binning, different NaN handling, whatever) and hit Run; the new result drops straight back into the conversation.

Yes. Because the analysis runs in your browser, there’s a practical ceiling of about 300 MB per file — roughly a few million rows. Files above that are blocked with a clear message rather than crashing the tab. Most real-world analyst CSVs are comfortably under this. (If you regularly work with larger data, let us know — server-side execution for big files is on the roadmap.)

CSV, Excel (.xlsx/.xls — every sheet becomes its own table), JSON (arrays of records), TSV, and Google Sheets. You can upload several files at once.

Yes. Load several tables — or a multi-tab spreadsheet — and ask a question that spans them. Curator detects the shared keys and writes the join for you, so “revenue by region, customers joined to orders” is one sentence.

Partly. The analysis executes locally in your browser, and we don’t keep a stored copy of your full dataset on our servers. But Curator is not an end-to-end-private tool: to turn your questions into working code, we send your column names and dtypes (and your results, with values masked) to a third-party LLM provider. Your rows aren’t sent, and when code fails and is auto-repaired only the exception type is forwarded — never the error message or any data value. If your data is sensitive, treat anything you analyze here as data you’ve shared with that provider. Full detail is in our Privacy Policy.

Yes, locally. Your notebooks and the data you upload are cached in your browser (localStorage + IndexedDB), so a notebook from days ago still has its data. It’s per-browser and per-device — open the same notebook elsewhere and you’ll be asked to re-upload. Nothing is synced to a server.

That’s the whole point. A chat answer is a guess; a chart that ships with the exact code, the assumptions, the trace, and the row counts is something you can verify. The code you read is the code that ran — what you check is what produced your numbers.

Open “Show your work” to see exactly how it was computed — the reasoning and the exact code, with its assumptions — then edit the code and re-run. You’re never stuck with a black-box result.

Curator runs DeepSeek (via OpenRouter) for code generation. You don’t pick or manage a model — that’s handled for you.

Free gives you 30 questions a month with every feature. Pro is $25/month for unlimited questions. No credit card needed to start — you upgrade from inside the app once you’re signed in.

Each analytical query (one that runs code to produce a result) counts toward the monthly limit. Conversational replies and re-running edited code don’t count.

Yes. Manage or cancel your subscription anytime from inside the app; access continues to the end of the paid period.