Cellprose
Give a model the whole spreadsheet, formulas included, and it answers right in one call.
Created by Utkarsh Gaikwad.
Cellprose (.sheet.md, or .cellprose.md) is an open plain-text format for Excel workbooks. It keeps what CSV and
MarkItDown drop: formulas, cell addresses, merged cells, number formats and named ranges. Every file is still valid
GitHub-Flavored Markdown, so workbooks render as tables and diff cleanly on GitHub.
What the benchmark showed
A pre-registered run on 11 Nov 2026: protocol, code and data frozen and the digest posted publicly before the first call. Four models (Claude Haiku 5.5 and Sonnet 5.5, GPT-6 Luna and GPT-6.1 Sol), 3 repeats, every token priced as uncached. Full tables: benchmarks/RESULTS.md.
- Accuracy in one call. On 30 real-world workbooks (163 questions), the workbook sent once as Cellprose got 93.8–99.3% right across the four models, against 65.9–84.1% for CSV and 70.0–85.7% for MarkItDown. The gap is in questions about formulas (trace, impact, hardcoded numbers), which CSV and MarkItDown cannot carry. Limit: the questions are generated by rule from the workbooks, not written by their authors.
- Cheaper than MarkItDown. Cellprose cost 0.32–0.60 times as much per correct answer, a win in 7 of 8 model-and-corpus pairs; the eighth (Sonnet on SpreadsheetBench) is no difference under the 1-point accuracy rule. Limit: 30 and 28 workbooks, so intervals are wide, Haiku's widest.
- Against CSV, wins for Sonnet and Luna on the real-world set and for Haiku on SpreadsheetBench; the rest are no difference. Limit: on SpreadsheetBench's value questions CSV already scores 83.5–99.7%.
- Against an agent that opens the
.xlsxwith a shell, accuracy is within a few points on most models (up to 3.8 points lower for Cellprose on some), but the agent takes 3.5–4.6 turns per question. Cellprose wins on cost for Sonnet on the real-world set; the other 14 of 16 comparisons are no difference. Limit: one-shot and agent loops spend tokens differently; this compares what a user pays. - No comparison was a loss. Run with Cellprose's own tools instead of a shell, an agent cost 0.29–0.79 times as much per correct answer as the shell agents, winning 14 of 24 comparisons (three overlapping views of the data, so the real-world set is counted twice) (Addendum A). Limit: that addendum was decided after the main results were seen; it was frozen and posted before its own first call.
Limits: 30 real-world and 28 SpreadsheetBench workbooks, 3 repeats, four models on one date, rule-generated questions, list prices checked 2026-11-11 with every token uncached; nothing pooled across models or corpora. Full list: benchmarks/RESULTS.md.
What it looks like
# Sheet: P&L
## A1:F4
#|A|B|C|D|E|F
-|-|-|-|-|-|-
1||Q1|Q2|Q3|Q4|Total
2|Units|1200|1296|1399.68|1511.65|5407.33
4|Revenue|54000|58320|62985.6|68024.4|243330
C2:E2 = B2*(1+Growth)
B4:E4 = B2*B3
Formulas are written once per fill range, not once per cell, and every value keeps its address. The spec is CC-BY-4.0, with a conformance kit for other implementations.
Use it
pip install "cellprose[mcp]" # PyPI release pending: until then, pip install -e ".[mcp]" from a clone
cellprose encode model.xlsx # -> model.sheet.md
cellprose decode model.sheet.md # -> model.xlsx (values, formulas, formats, merges, names)
cellprose stats model.xlsx # chars and tokens vs CSV and Markdown (needs the [stats] extra)
import cellprose
text = cellprose.dumps("model.xlsx") # put this in the prompt
There are two ways to use it, in the prompt or with its own tools; the guide compares them.
In an agent. claude mcp add cellprose -- cellprose-mcp gives an agent tools to outline a workbook, read ranges,
find cells, trace a formula and run a what-if (index_workbook, read_ranges, find_cells, trace_cell, whatif, write_cells), plus an agent skill in skills/cellprose/. A hosted server at
tablient.com does the same without installing anything. See INTEGRATIONS.md.
Where you already load spreadsheets (integrations/, publishing with the package):
| You use | Swap in |
|---|---|
| MarkItDown | pip install markitdown-cellprose (not yet on PyPI), then markitdown --use-plugins book.xlsx |
| LangChain | from langchain_cellprose import CellproseLoader and CellproseLoader("book.xlsx").load() |
| LlamaIndex | SimpleDirectoryReader("data", file_extractor={".xlsx": CellproseReader()}) |
Formula cells with no cached value print as ? unless LibreOffice is installed (recalc="auto"). A decoded
.xlsx has no cached values until Excel or LibreOffice opens it.
Status: pre-release, v0.3. Encode, decode, recalculate and encode again gives byte-identical text on the benchmark workbooks and an edge-case set (E14).
Cellprose was called Sheetmark until October 2026. Files with the old
sheetmark:front-matter key still read. Thesheetmarkname on PyPI belongs to an unrelated project; installcellprose.
Repository layout
| Folder | What's in it |
|---|---|
src/cellprose/ |
The package: writer, reader, CLI, MCP server |
integrations/ |
MarkItDown plugin, LangChain loader, LlamaIndex reader |
skills/cellprose/ |
Agent skill with a bundled converter |
docs/ |
Design doc, spec, benchmark protocol, git guide |
benchmarks/ |
Workbook sets, encoders, runners, scorer, all raw results |
experiments/ |
One file per experiment, linked to the benchmark code and results |
conformance/, js/, viewer/, site/ |
Conformance kit, JS writer, GitHub viewer, website |
tests/ |
Python tests |
Sponsor
Cellprose is free and open source. If it saves you time or tokens, you can sponsor its development on GitHub Sponsors.
Licence
Code, website and benchmark tooling are under Apache-2.0. The format specification (docs/SPEC.md) and the conformance kit are under CC-BY-4.0. The name "Cellprose" is not licensed by either.
Metadata
Release files for cellprose 0.3.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| cellprose-0.3.0.tar.gz | 44.7 kB | With provenance | Details for cellprose-0.3.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| cellprose-0.3.0-py3-none-any.whl | Python 3 | none | any | Details for cellprose-0.3.0-py3-none-any.whl |
Total release size: 88.4 kB
Release files / cellprose-0.3.0.tar.gz
| Download URL | cellprose-0.3.0.tar.gz |
|---|---|
| Size | 44.7 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
0130b940d8d6820857e7bf17d39ff92e26e8fcf18cc3bd53a7d3ba5d9dde5500
|
|
BLAKE2b-256 checksum How to use checksums |
8f3760b8bb023e0349bd97ab00a48ea741ff57384397d2ca2c42c52ad23a0ef3
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.
Transparency logRelease files / cellprose-0.3.0-py3-none-any.whl
| Download URL | cellprose-0.3.0-py3-none-any.whl |
|---|---|
| Size | 43.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
5ebe4c05ae61ed01c0cc995954fca0dcbde45f8d8a08638492360888b1287e8e
|
|
BLAKE2b-256 checksum How to use checksums |
8a6bb17b63ec95e590ee98e19d59911d93a563306802df618db462a87271344f
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.
Transparency logSecurity for cellprose 0.3.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for cellprose-0.3.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.