Curator

Privacy Policy

Last updated: June 5, 2026

Short version

Curator runs the analysis as Python in your browser, but it is not a fully private or local-only tool. To turn your questions into working code — and to correct it when a first attempt fails — we send information about your data to a third-party LLM provider (OpenRouter / DeepSeek). That includes your questions and your dataset's structure (column names and types). Your rows are not sent, and the values in your results are masked before they're sent and restored in your browser. When generated code fails, only the error TYPE (e.g. "ValueError") is sent to repair it — never the message or any value. If your dataset is sensitive, treat anything you analyze here as data you've shared with that provider.

What we collect

  • Account info: the email address and (if you sign in with Google) display name and avatar URL associated with the account you create through Firebase Authentication.
  • Query & data content: the questions you ask, plus your dataset's structure — column names and dtypes — and the results your analysis produces, with the actual values masked. Your rows are not sent. This content is forwarded to our LLM provider to generate the analysis code. When code fails and is auto-repaired, only the exception TYPE (e.g. "KeyError") is sent — never the error message or any data value.
  • Operational logs: error reports and basic usage telemetry so we can debug failures. These may contain fragments of the data involved in an error.
  • Query traces: to find and fix slow or failing analyses, we record a timing trace for each query — your question, your dataset's structure (column names, types, row count), and how long each step took. This is retained on our servers for product improvement. It never includes your cell values or row contents.

How your data is handled

  • The analysis itself executes in your browser via Pyodide. We do not keep a stored copy of your full dataset on our servers — it lives in your browser session.
  • However, Curator is not end-to-end private: portions of your data are transmitted to the LLM provider as described above in order to generate and refine code.
  • Passwords. Firebase handles authentication for us.
  • We don't sell or share your data with advertisers, brokers, or any third party for marketing purposes.

Improving the shared library (optional, off by default)

Curator ships a shared library of reusable analysis methods. You can opt in — from the account menu — to help it grow. This setting is off unless you turn it on.

  • What's shared when you opt in: when you run an analysis, your browser sends the anonymized structure of the code it ran — every column reference and literal value is first replaced, in your browser, with a placeholder (col_0, col_1, …), so what leaves is a skeleton like df[col_0].quantile(p), never df["salary"].quantile(0.9). We also send the result type (e.g. “a single number”) and the operation involved.
  • What is never shared: your questions, your column names, your row values, your results, or the specific arguments a method was run with. If the structure can't be fully anonymized, that analysis is dropped rather than sent. The sensitive parts never leave your browser — we capture the reusable pattern, not your data.
  • How it's used: these structural skeletons are gathered, and a pattern is only ever looked at once it has recurred across many different people. A common pattern is then turned into a clean, reusable method.
  • Nothing is published automatically. Every candidate method is screened for anything that looks like identifying information and reviewed by a person before it could ever be added to the shared library. The shared library is never grown directly from anyone's data.
  • You can turn this off at any time in the account menu; doing so stops any further sharing immediately.

Third parties

  • Firebase (Google): account management.
  • OpenRouter / DeepSeek: the LLM provider that receives your questions, dataset structure, and analysis results (with values masked) to generate and refine code. Your raw data values are not sent — not even in error messages during automatic code-fixing, where only the exception type is forwarded.
  • Vercel: hosts our application.

Your rights

You can delete your account at any time by contacting us at support@askcurator.io. We'll remove your account data within 30 days.

Contact

Questions, requests, complaints: support@askcurator.io.