Compare
The best DOCX MCP servers for Microsoft Word
There are now a dozen MCP servers that let agents edit Word documents. We ran four of them, plus Anthropic's DOCX skill and a plain python-docx harness, on the same 279 editing tasks and measured pass rate, cost and latency.
Benchmark
Results on 279 Word editing tasks
| Solution (GPT 5.6 Sol) | Pass rate | Median cost | Median latency | 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 |
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.
By task type
Where the MCP servers differ
The gaps are widest on tables, headings and tracked changes. Those edits depend on inherited styles and document structure, which is where tool-by-tool OOXML editing breaks down.
| 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% |
The servers
What each one is
Vespper DOCX MCP
Hosted MCP with three tools: read, edit and search. The agent reads the document as clean HTML and sends back the HTML it wants. Vespper writes the OOXML and returns the edits as Word tracked changes.
Office CLI MCP
A general-purpose Office toolkit exposed over MCP. It covers a lot of ground, but in our runs agents needed about 13 tool calls per task, which made it the slowest and most expensive option.
Adeu MCP
An MCP server for redlining Word documents. It was cheap and quick per task, but passed fewer than half of the tasks.
SuperDoc MCP
The MCP server from SuperDoc, an open-source DOCX editor with SDKs.
Not benchmarked
Other Microsoft 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.
safe-docx (Junior)
TypeScript MCP server focused on text replacement and redlines in existing documents.
docx-mcp (SecurityRonin)
Open-source server for reading and editing .docx files with tracked changes and comments.
Microsoft Word MCP (Work IQ)
Microsoft's preview server in Agent 365 for creating and reading Word documents in OneDrive and SharePoint.
Use cases
Headless, in your editor, or inside Word
Headless
Run agents in your backend with no editor at all. Connect to one hosted endpoint, with no sandbox to run. Create a session once, and the agent gets back the .docx with its edits as Word tracked changes.
Your editor
Vespper is editor-agnostic, so it works with any editor you already use. Edit streaming sends incremental updates as the agent writes, so users watch edits appear live.
Word add-in
Bring the agent into Microsoft Word itself. Our open-source sample add-in combines streaming and document caching, so follow-up tool calls take about 400ms even on large documents.
Setup
Connect Vespper DOCX MCP
One hosted endpoint, no infrastructure. The quickstart covers authentication and file handling.
from fastmcp import Client
client = Client(
"https://mcp.vespper.com/mcp",
auth="TOKEN",
)
async with client:
tools = await client.list_tools()
# Three tools: read, edit and searchRun your own documents through it.
The free plan includes 500 edits a month. See pricing.