Ask your spreadsheet a question

Drop in a CSV, a spreadsheet or a PDF table. It checks the data for problems first, writes the code, runs it, and answers with charts you keep.

Analyse a spreadsheetFree to try. Nothing to install.

How it works

Three steps, and none of them is “configure”.

01
Give it the data
Drop a file anywhere on the page, or start with the sample dataset that is already loaded. CSV, spreadsheets and PDF tables all work.
02
It profiles before it counts
Shape, column types, null counts, duplicate rows, dates in three formats in one column — it finds these and tells you before drawing a single conclusion from them.
03
You get the answer and the working
The finding in plain language, the chart as a PNG in your workspace, and the Python that produced both. Nothing is described that was not actually run.

What it can do

Real analysis, not an impression
pandas in a Linux sandbox. It writes a script, runs it, reads the output, and fixes the script when the output is wrong — the same loop you would run yourself.
Charts you can keep
PNGs written to your workspace with titles, axis labels and units. It looks at the image it drew and redraws it if the labels overlap, because a chart that saved successfully can still be unreadable.
It tells you what is wrong with your data
Numbers stored as text, thousands separators inside quoted values, duplicated headers halfway down — the things that make a total silently wrong. Ask what is wrong before you ask for a total.

How it is set up

The mechanics, so you know what you are getting before you sign in.

Environment: A Linux container with Python 3, pandas and matplotlib already installed. It starts in your workspace and keeps its state between your messages.
Tools: Read, write and edit files, glob and grep, a shell, and the ability to look at an image it drew or read a PDF you gave it.
Model: Any chat model in the catalogue, changed mid-conversation. The analysis is real code either way — the model decides what to run, not what the answer is.
Starting files: A deliberately imperfect sample sales dataset, so the profiling step has something real to catch on the very first run.
What persists: Files, installed packages and shell state survive between messages. Charts and scripts are still in your workspace when you come back to the session.

Things people ask it

“Which region grew fastest, and show me a chart of it.”
“Before I trust any totals — what is wrong with this data?”
“Join these two files on order id and tell me what did not match.”
“Plot revenue by month and mark anything more than two standard deviations out.”

What it will not do

Every one of these is a real constraint we have hit, not a roadmap item.

It will not invent a column. If your question needs data that is not in the file, it says so rather than approximating an answer.
Uploads still pass through your browser, so there is a size ceiling. Genuinely large tables are the next thing we are fixing.
It reads what you give it. It has no access to your database, your drive, or anything you have not put in the workspace.

Which model does this best

Measured over real runs of this agent, per model. A run counts as finished when it produced the answer on its own — nothing failed, nobody was asked to approve anything, and it did not run out of steps. Read all three columns together: a model that finishes fast by giving up scores badly on the first, and one that completes everything by grinding scores badly on the second.

Model	Completion rate	Median cost	Median time	Runs
gemini-3.5-flash-litegoogle	100%	$0.0077	27s	5
claude-sonnet-5anthropic	100%	$0.061	55s	3

Completion rate: Runs that finished the job without you having to step in.
Median cost: What a typical run costs, in credits.
Median time: Wall clock, from the first message to the answer.

Medians over the runs behind each row. A model appears once it has 3 runs on this agent, and the run count is shown so you can judge how much a figure rests on.