Compare DOCX MCP servers for Word editing agents
Choose a server for editing existing DOCX files and returning changes users can review. Compare deployment, document handling, and native revisions alongside Vespper's results on 279 editing tasks.
Integration
What your application needs to supply
Start with where the DOCX can be processed, how your application retrieves the edited file, and whether the output must contain native Word revisions. A server appearing in an MCP client's tool list does not establish that upload, editing, and delivery work.
Documentation checked . These are documented integration requirements; the benchmark below reports the separately listed tested versions.
| Server and sources | Runtime and access | Document workflow | Review output and limits |
|---|---|---|---|
| Vespper DOCX MCP Server | Hosted Streamable HTTP with a bearer API key. The DOCX is uploaded to a document session. | Upload once, attach the session ID to read/search/edit calls, retrieve the latest DOCX bytes, and close the session. | New edits are tracked by default, with configurable author and tracking. |
| Office CLI MCP | Local OfficeCLI process over stdio. Your host supplies access to document files. | Pass CLI commands through the officecli tool to read and modify files. Your application manages paths and output delivery. | Explicit revision properties create tracked changes or accept/reject revisions. Move tracking is limited to runs. |
| Adeu MCP | Local Node.js or Python MCP process. The host must have access to the input and output files. | Read a Markdown/CriticMarkup representation, apply targeted batches, and write the edited DOCX. | Native tracked changes and comments. Its text projection omits some layout properties and treats images as read-only. |
| SuperDoc MCP | Local stdio process launched by the MCP client. Current setup requires Node.js 20 or newer. | Open a local path into an in-memory session, inspect and edit it, save a copy, then close the session. | Tracked replacements use an explicit change mode. Inspect receipts and the saved DOCX; saving overwrites an existing output path. |
Local processing describes where the document engine runs. An agent can still send text from tool results to its configured model provider. Include that path when choosing your deployment and data-handling model.
Vespper-run benchmark
Results on 279 Word editing tasks
| Solution (GPT 5.6 Sol) | Pass rate | Median cost | Median latency | Median tool calls |
|---|---|---|---|---|
| Vespper DOCX MCP | 81.4% | $0.054 | 14.6s | 3 |
| DOCX skill (Anthropic) | 78.1% | $0.157 | 51.5s | 10 |
| python-docx harness | 77.4% | $0.112 | 33.4s | 4 |
| Office CLI MCP | 65.2% | $0.240 | 57.1s | 13 |
| Adeu MCP | 44.8% | $0.078 | 24.0s | 5 |
| SuperDoc MCP | 41.7% | $0.130 | 16.8s | 4 |
We evaluated four MCP servers, Anthropic's DOCX skill, and a python-docx harness on the same internal test set. A pass means the edited blocks matched the reference content and rendered styles under our published scoring method. These results describe that workload and configuration.
Model: GPT 5.6 Sol at medium reasoning, same system prompt for every solution. Setup steps like opening, saving and loading a skill were handled for every candidate. Versions: SuperDoc MCP v0.18.1, Office CLI v1.0.145, Adeu MCP v3.0.2. The full methodology covers the dataset, scoring, and GPT 5.6 Terra results. The setup steps were handled outside the agent loop for every candidate, so these timings do not measure installation or the full upload-to-download experience.
By task type
Where the MCP servers differ
Use the task groups to choose what to test on your own documents. A task can belong to more than one group, so the counts overlap. A zero pass rate in a tested configuration does not mean the current product has no tools for that operation.
| Task type (GPT 5.6 Sol pass rate) | Vespper | Office CLI | Adeu | SuperDoc |
|---|---|---|---|---|
| Tables (n=90) | 77.8% | 54.4% | 27.8% | 25.8% |
| Lists (n=96) | 71.9% | 51.0% | 45.8% | 34.4% |
| Paragraphs (n=159) | 75.5% | 60.4% | 40.3% | 39.6% |
| Headings (n=28) | 60.7% | 35.7% | 21.4% | 17.9% |
| Tracked-change management (n=26) | 57.7% | 34.6% | 34.6% | 0.0% |
Not benchmarked
Other Word MCP servers
These weren't part of our benchmark, so we have no numbers for them. Descriptions are from each project's own documentation.
Office-Word-MCP-Server (GongRzhe)
Open-source Python server for creating, reading and formatting Word documents.
Official documentationsafe-docx (Junior)
Local MCP server and CLI for reading, editing, comparing, and saving documents. Its documentation describes clean and tracked-changes output.
Official documentationdocx-mcp (SecurityRonin)
Open-source server for reading and editing .docx files with tracked changes and comments.
Official documentationMicrosoft Word MCP (Work IQ)
Microsoft's preview Word server creates and retrieves documents in OneDrive and SharePoint and adds or replies to comments. Its listed tools do not document body-text replacement or native redlining.
Official documentation
Use cases
Headless, in your editor, or inside Word
Headless
Run the agent in your backend. Your application opens the document session, retains the latest DOCX, and returns a download or attachment.
Build a document-editing agentYour editor
Vespper is editor-agnostic. Your application applies returned DOCX revisions to its editor and handles revision ordering and simultaneous user changes.
See an editor integration exampleWord add-in
Use the open-source Office.js add-in to read the active document, stream edits back into Word, and follow Word's tracking preference.
Build the Word add-inFirst edit
Upload, edit, and return a DOCX
The quickstart includes a sample DOCX and complete Python and TypeScript integrations. Your application supplies a Vespper API key and its agent's model configuration. The DOCX MCP Server exposes three tools: read_document, search_document, and edit_document.
- Download the quickstart's sample and run the integration for your agent framework.
- Open a document session, wire the three tools with its session metadata, and ask for the sample edit. Use the returned HTML when constructing edit anchors.
- Save the latest DOCX before closing the session. Check applied and failed edits; successful parts of a batch may already have changed the document.
- Open the saved file in Word. Confirm the requested change, author, and accept/reject behavior, then repeat with a document representative of your application.
For review workflows, include a separate pending revision and confirm it survives your edit. The tracked-changes reference explains author settings and selective revision handling. Check the edit API and published limitations against the operations your users need.
Run your own documents through it.
The free plan starts you with 500 edits. See pricing.