Skip to content

Why Cellprose ​

Models read spreadsheets badly when they only see values. CSV hides which numbers are formulas. The raw .xlsx buries them in XML spread over a dozen files. Record formats like TOON drop row and column positions, so a model can no longer say which cell it means.

Cellprose keeps the grid with its addresses, and puts the logic in a short block of fill ranges after it. A model reads what the sheet does instead of reverse-engineering it from numbers, and in the benchmark that showed up as accuracy: on 30 real-world workbooks, one call with Cellprose answered 93.8–99.3% of questions correctly, against 65.9–84.1% with CSV, mostly on questions that need a formula. See the benchmarks for the verdicts, intervals and limitations.

Every Cellprose file is valid GitHub-Flavored Markdown, so a workbook renders as tables on GitHub and a change to one cell is a one-line diff.

markdown
# Sheet: Forecast

## A1:N5

|#|A|B|C|D|E|F|…
|-|-|-|-|-|-|-|…
|2|Bookings|100000|104000|108160|112486|117000|…

```formulas
C2:E2 = B2*(1+$B$7)
G2:I2 = F2*(1+$B$7)
```

F2 is in neither range: it holds a typed number where a formula belongs. That gap is visible in two lines of text; in CSV it is invisible.

Two ways to use it ​

In the promptWith its own tools
Howcellprose encode book.xlsx or cellprose.dumps(...) writes the workbook as text, and you put it in the message.cellprose-mcp gives an agent tools: index_workbook, read_ranges, find_cells, trace_cell, whatif. The model asks for the cells it needs.
CallsOne.Several turns.
SuitsA workbook that fits in the model's context; a one-shot answer in a pipeline.A workbook too big to send whole; tracing a formula, trying a what-if, editing. This is guidance, not something the benchmark tested.
Measured, 11 Nov 2026 run93.8–99.3% correct on 30 real-world workbooks, against 65.9–84.1% for CSV and 70.0–85.7% for MarkItDown. Cost per correct answer 0.32–0.60 times MarkItDown's.Cost per correct answer 0.29–0.79 times that of agents with a shell (Addendum A, 14 of 24 comparisons, three overlapping views). That addendum was decided after the main results were seen.

Both read the same text. The first sends it all at once; the second lets the model look things up. See the benchmarks for the numbers and their limits.

Quick start ​

From a checkout of the repository, until the first package release:

bash
pip install -e ".[dev]"
cellprose encode model.xlsx          # writes model.sheet.md
cellprose decode model.sheet.md      # back to .xlsx: values, formulas, formats, merges, names
cellprose check model.sheet.md       # validate without writing anything
cellprose stats model.xlsx           # size as Cellprose vs CSV and Markdown

For agents, there is an MCP server and an agent skill:

bash
pip install -e ".[mcp]"
claude mcp add cellprose -- cellprose-mcp

Or try it without installing anything: the playground converts an .xlsx in your browser.

What it keeps and what it drops ​

Cellprose keeps values, formulas, number formats, merges, defined names and sheet structure, and converts back to .xlsx without losing them. It drops styling (fonts, fills, borders), charts, pivot tables, conditional formatting, data validation, macros, comments and images. The specification lists the limits.

Cellprose is an open format for spreadsheet workbooks. Benchmarks: pre-registered run of 11 Nov 2026, 58 held-out workbooks, four models, three repeats. Privacy

CellproseGitHub ↗
A1fx
Ready