Wolf Workspace / Word / Excel / PowerPoint

# Your team and AI agents, *editing Office files together.*

An agent proposes the edit. You see what would change and approve it before it saves. Wolf Workspace keeps every revision, and the Office kernel underneath changes only what was asked.

[Start a 14-day free trial](https://workspace.wolfsdk.com/-/sign-up)[See how Wolf Workspace works](https://wolfsdk.com/workspace)

No card needed. Team is $10 per seat each month after the trial, and AI agents are $0. Or self-host the open-source workspace.

Hosted workspace / Built-in assistant / MCP

## Ask for the edit. *Approve it before it saves.*

Wolf Cloud is a shared workspace at workspace.wolfsdk.com for Word, Excel, PowerPoint, and PDF files. Open a Word document, tell the built-in assistant what to change in plain language, and it proposes the edit for you to approve or reject, with a receipt of the parts it changed. You need no agent and no model key; teams with their own AI agents connect them over MCP.

14-day free trial with no card. Team is $10/seat/month ($100/seat/year, from 1 seat), Business is $20/seat/month ($200/seat/year, from 2 seats, OIDC included), the built-in assistant includes 300 requests per seat each month, and AI agents acting for a signed-in user are $0. [Terms](https://wolfsdk.com/terms) · [Refund Policy](https://wolfsdk.com/refund-policy) · [Privacy Notice](https://wolfsdk.com/privacy)

[See Wolf Cloud pricing](https://wolfsdk.com/pricing)[Open workspace.wolfsdk.com](https://workspace.wolfsdk.com/)[Explore Wolf Workspace](https://wolfsdk.com/workspace)

Built on the Office kernel

## Why an agent's edit is *safe to approve.*

Other libraries rebuild your file, and every one we tested broke something. The Wolf engines under Wolf Workspace edit the file in place and change only what you asked. The same engines install on their own for developers.

Same edit, two libraries

## Change one cell. *Keep the whole workbook.*

We changed APAC's Q3 forecast from 4.1 to 4.4 in a workbook saved by Excel, once with openpyxl and once with Wolf, then opened both results in Excel.

openpyxl4 of 4 features lost

![The forecast workbook in Excel after the openpyxl edit. Q3 for APAC reads 4.4 and its total 16.5, but the Trend column is empty, the yellow DRAFT note is gone, the Total data bars are faded gradients, and the chart lost its style: a smaller title and rounded corners.](https://wolfsdk.com/examples/edit-in-place/compare/excel-openpyxl.webp)

14 of 17 parts rewritten, 2 removed

- Trend sparklinesLost
- DRAFT sticky noteLost
- Solid data barsLost
- Chart styleLost

Wolf4 of 4 features kept

![The forecast workbook in Excel after the Wolf edit. Q3 for APAC reads 4.4 and its total 16.5. The Trend sparklines, the yellow DRAFT note, the solid data bars and the chart match the original.](https://wolfsdk.com/examples/edit-in-place/compare/excel-wolf.webp)

1 of 17 parts rewritten, 16 untouched

- Trend sparklinesKept
- DRAFT sticky noteKept
- Solid data barsKept
- Chart styleKept

Check it yourself

- [Original workbook](https://wolfsdk.com/examples/edit-in-place/compare/q3-forecast.xlsx)
- [openpyxl result](https://wolfsdk.com/examples/edit-in-place/compare/q3-forecast-openpyxl.xlsx)
- [Wolf result](https://wolfsdk.com/examples/edit-in-place/compare/q3-forecast-wolf.xlsx)
- [openpyxl's warnings](https://wolfsdk.com/examples/edit-in-place/compare/openpyxl-warnings.txt)
- [results.json](https://wolfsdk.com/examples/edit-in-place/compare/results.json)
- [SHA256SUMS](https://wolfsdk.com/examples/edit-in-place/compare/SHA256SUMS)

Reproducible examples

## Three runs *you can repeat.*

Each example comes with its script, inputs or links, receipts and checksums, so you can rerun it against the published wolf binary.

Spreadsheet to PDF

### One soffice command, two IRRs.

Convert a public financial model with headless LibreOffice and read a 44,884.78% IRR. Run the same command with Wolf and read NM, the value shown in Excel.

[See the two PDFs](https://wolfsdk.com/examples/impossible-irr)

Extraction

### Hidden text, labelled for the agent.

Plant instructions in hidden runs, comments, notes and very hidden sheets. Run wolf extract to label each channel with visible: false and a reason.

[Compare the three views](https://wolfsdk.com/examples/hidden-text)

Transactions

### A broken edit, refused.

Ask for an edit that breaks a ledger tie-out and get exit 7, with the invariants checked in memory and the file left byte-identical. Resend the repaired request to commit it.

[Watch the refusal](https://wolfsdk.com/examples/refused-edit)

Choose your document job

## Familiar formats. *Explicit coverage.*

Each engine page covers supported operations, licensing and installation.

XLSX and XLSM

### WolfXL

Edit supported .xlsx and .xlsm workbooks without rewriting the package.

The community edition is MIT, on public PyPI. Commercial editions are licensed separately.

[WolfXL product page](https://wolfxl.com/)

PPTX

### WolfPPT

Update text, tables and chart data in existing .pptx decks.

The MIT community edition is the place to start. It is free, including commercial use. WolfPPT Commercial adds rendering, assembly and more formats as a subscription for one legal organization.

[WolfPPT product page](https://wolfppt.com/)

DOCX

### WolfDocx

Update table cells, replace text, fill template rows and accept tracked changes in existing .docx files.

The MIT community edition is on public PyPI and free, including commercial use. WolfDocx Commercial is the full engine, as a subscription for one legal organization.

[WolfDocx product page](https://wolfdocx.com/)

PDF

### WolfPDF

Fill forms, redact, sign, compose and compare existing PDFs, with every result verified before it is saved.

Commercial only. Ask about licensing for your integration.

[Wolf Runtime](https://wolfsdk.com/runtime)

Install wolf 0.2.25

```
curl -fsSL https://wolfsdk.com/install | sh
```

Runs on macOS arm64 and Linux x64 with no account, key or credential. Windows x64 is not yet supported by the public installer. Free for individual and non-commercial use under the [Wolf Free License](https://wolfsdk.com/download/wolf/v0.2.25/LICENSE-WOLF-FREE.txt).

Business use is not covered by the Wolf Free License; get a 60-day nonproduction evaluation at [https://wolfxl.com/evaluate](https://wolfxl.com/evaluate).

Then, for Claude Code, Cursor or Codex, run `wolf install` to add the wolf MCP server and its editing skill to each one it finds, on macOS and Linux. Read the [MCP setup](https://wolfsdk.com/runtime/docs#integrations).

WolfRender renders for the document engines. The command-line tool defaults to bundled fonts, and render operations report their omissions. [About WolfRender](https://wolfrender.com/)

Measured on 138 test files

## No unrelated *part changed.*

In the wolf-v0.2.0 preservation snapshot, saving the same files with LibreOffice 26.2.5.2, our control, changed 5,904 parts.

| Format | Test files | Wolf, no-op save | Wolf, edit | LibreOffice, no-op save |
| --- | --- | --- | --- | --- |
| XLSX | 45 | No unrelated part changed | wolf-v0.2.0: recalculation save in the 44 files that finished: no unrelated part changed | 1,195 part changes |
| DOCX | 55 | 55 of 55 byte-identical | Tracked changes accepted in 52 files: no unrelated part changed | 1,400 part changes |
| PPTX | 38 | 38 of 38 byte-identical | Text edit in the 37 files that contain text: no unrelated part changed | 3,309 part changes |
| Total | 138 |  |  | 5,904 part changes |

Historical wolf-v0.2.0 preservation snapshot. Same-format operations on test files. LibreOffice 26.2.5.2 is the control.

Read more

- [How preservation is tested](https://wolfxl.com/openpyxl-preservation)
- [Receipts in practice](https://wolfsdk.com/blog/introducing-wolf-runtime)

[Read the held-out DOCX agent-edit benchmark](https://wolfsdk.com/agent-edit-benchmark)

Measured against Office output

## Closer to Office's output *than LibreOffice.*

In the wolf-v0.2.21 heldout-2026-10f receipt, Wolf has lower mean distance to Office's own PDFs than LibreOffice in DOCX, PPTX and XLSX.

| Format | Documents | Wolf mean distance | LibreOffice mean distance | Wolf page counts match |
| --- | --- | --- | --- | --- |
| DOCX | 42 | 0.088 | 0.203 | 39/42 |
| PPTX | 51 | 0.087 | 0.120 | 51/51 |
| XLSX | 38 | 0.161 | 0.335 | 31/38 |

Mean SSIM distance to Office's own PDF output on a frozen held-out corpus.

wolf-v0.2.21, heldout-2026-10f, November 2, 2026: mean SSIM distance to Office's own PDF output, rounded to three decimals; lower is closer. The last column counts matching page counts. Receipt: evidence/ltg-l4-heldout-checkpoint-10f.json. The paired cohort has 42 DOCX, 51 PPTX and 38 XLSX documents; one XLSX native baseline is missing. There are 10 document losses against LibreOffice, 23 per-document SSIM regressions and a retained uk12 pagination regression. Scoring-helper identity and local-only declaration ordering remain unqualified. Native Office oracles were reused; this is a regression benchmark, not a blind test. Omission counts from older passes are not requalified, and no PDF-input ranking exists.

[Read the evidence and limitations](https://wolfsdk.com/runtime#rendering)[Compare Wolf, BetterOffice, office2pdf and LibreOffice](https://wolfsdk.com/blog/office-rendering-comparison)

Who it's for

## Built for whoever *signs off on the edit.*

Agent builders

### Let agents touch real files.

Give reviewers a diff they can approve in seconds, and keep the receipt as an audit trail or a reward signal.

Enterprise teams

### Show what didn't change.

Hand compliance a hash for every part outside the edit, so they can check the rest of the file for themselves.

Python developers

### Save without the cleanup.

Keep the charts, images and extensions a load-and-save round trip can drop, with no manual fix-up pass afterward.

Sandboxes / Platforms / Embedded products

## Run the Office kernel *inside your environment.*

Wolf Runtime by WolfSDK embeds the kernel for XLSX, DOCX, PPTX and PDF behind one command and receipt contract.

Starter $4,999 a year, Internal from $15,000 a year, Product from $60,000 a year, Platform from $150,000 a year. Starter is bought online by small organizations; Internal, Product and Platform need a written Order.

[Explore Wolf Runtime](https://wolfsdk.com/runtime)

## Run it on the file *you trust least.*

Pick the messiest deck, workbook or document you have. Edit it, diff it and read the receipt, then choose an engine for your workflow or Wolf Runtime for an embedded deployment.

[See developer pricing](https://wolfxl.com/pricing)[Explore Wolf Runtime](https://wolfsdk.com/runtime)

**Notes on the measurements (8)**

1. Library comparison: one workbook built with XlsxWriter, saved by Microsoft Excel 16.113.3 for Mac, with the author and save folder blanked afterward. The workbook was built to carry sparklines, a data bar, a text box, a chart, a note and list validation, features common in Excel-authored files. The same cell edit was made with openpyxl 3.1.5 (load_workbook, assign, save) and with a wolf 0.2.2 build (commit 3dc3bb2) on Linux x86_64; Wolf also recalculated the dependent total. All three files opened in Excel with no repair prompt; the screenshots are Excel's own window, cropped to the grid. openpyxl warned on load that it would remove the sparkline group, conditional formatting and unknown extensions. Results are for this file; openpyxl documents that it does not read all items in a file.
2. Demo deck: one slide with a plan versus actual column chart, generated with python-pptx 1.0.2 and saved by Microsoft PowerPoint 16.113.3 for Mac; author metadata was blanked afterward. The chart stores raw dollars and shows them in millions through the number format of its data, which its data labels and value axis are linked to. Both tools changed APAC's actual from 4,100,000 to 4,400,000. python-pptx ran the published script, which reads the chart's own categories and series and hands them back to chart.replace_data; replace_data writes General as the number format unless the script passes the original format back in, and this script does not. It rewrote 34 of 40 parts: the chart, a regenerated chart data workbook, 27 parts that match the original as canonical XML and 5 that hold the same entries in a different order. Wolf ran deck-refresh from wolfppt 0.2.0 built from source (commit 3dc3bb2) on Linux x86_64, with a request bound to the deck's SHA-256, the chart's identity and its current values; it changed one cached value in the chart and one cell in the chart's data workbook, keeping that cell's format, and rewrote 2 of 40 parts. PowerPoint opened all three decks and exported each to PDF; the images are those PDF pages, and every label the page reports was read back from those PDFs. The ring marks the APAC actual label across the three renders. ZIP compression bytes may differ after repacking; part contents do not. Hashes are in the published SHA256SUMS.
3. Word memo: generated with python-docx 1.2.0 and saved by Microsoft Word 16.113.3 for Mac, with the author fields blanked afterward. The memo was built to place Q2 in 8 places: a header, the title, a large plain-text content control, body text, a table heading, a hyperlink, a text box and the closing line. python-docx ran the published script, which edits run by run through body paragraphs, hyperlinks, tables and headers; its API has no access to content controls or text boxes, which code editing the raw XML could reach. It replaced 6 of 8. A wolf 0.2.2 build (commit 3dc3bb2) on Linux x86_64 ran replace-text with scope body+headers, which covers the main document and its headers but not footers, footnotes or comments, and leaves a match split across formatting runs unedited with a warning; every Q2 in this memo sat in one run. Wolf replaced 8 of 8 and rewrote 2 of 18 parts. Word opened all three files and exported each to PDF; the images are those PDF pages, cropped to the text. Rings mark the position of each Q2 word that PDF text extraction found in the original, and a ring is red where the word at that position still reads Q2 in a result. Of python-docx's 10 rewritten parts, 2 hold the edit, 7 match the original as canonical XML and 1 holds the same entries in a different order.
4. Preservation: same-format operations on finite sets of test files (45 XLSX, 55 DOCX, 38 PPTX), run on Linux x86_64 with the wolf 0.2.0 release binary, each run capped at 8 GiB of memory. One XLSX workbook exceeded that cap during recalculation and wrote no output, so the recalculation result covers 44 files. The DOCX edit is tracked-change acceptance on revisions the test injected; plain DOCX text replacement is outside this measurement. XLSX recalculation rewrites formula sheets and the calculation chain, which we count as intended. One image-only PPTX deck has no text and is excluded from the text-edit count. Part equality is not whole-archive equality after an edit. No claim is made for arbitrary edits or conversions. LibreOffice 26.2.5.2 is the control; an office suite that re-serializes on save is expected to rewrite parts, and it did so in every test file.
5. Rendering: wolf-v0.2.21, heldout-2026-10f, November 2, 2026, evidence/ltg-l4-heldout-checkpoint-10f.json. The paired cohort has 42 DOCX, 51 PPTX and 38 XLSX documents, scored as mean SSIM distance to Office's own PDF output (lower is closer) against a stored LibreOffice 26.8.0.3 baseline. One XLSX native baseline is missing. Ten document losses against LibreOffice, 23 per-document SSIM regressions and the uk12 pagination regression remain. Scoring-helper identity and local-only declaration ordering remain unqualified. Native oracles were reused, earlier failures were fixed between scoring passes, and no blind test has been run. Historical omission counts of 12 DOCX, 18 PPTX and 8 XLSX belong to the October 10 launch-article snapshot and are not requalified for this receipt. Equal page counts are not pixel equality. Rendering does not establish calculation correctness, and no PDF-input ranking exists.
6. Receipts and omissions: the command and receipt contract is documented, and operations differ by format. A receipt records what Wolf did and does not prove every semantic outcome. Omission reports cover the losses Wolf detects.
7. Local processing: the embedded binary has no network client, runs no subprocess, interprets no macros, resolves no external references and collects no telemetry. Optional recognition and session features run as separate processes. This is a design statement and carries no hostile-file or certification claim.
8. Fonts: the command-line tool defaults to bundled fonts. The drop-in command shims default to system fonts.
