Compare
Vespper DOCX MCP vs the Anthropic DOCX skill
Both let an agent edit Word documents. On our benchmark of 279 editing tasks, Vespper was 2.7–2.9× cheaper and 2.7–3.5× faster than the DOCX skill, and more accurate.
Benchmark
The headline numbers
| Metric (GPT 5.6 Sol) | Vespper DOCX MCP | DOCX skill |
|---|---|---|
| Cost per task | $0.054 | $0.157 |
| Task accuracy | 81.4% | 78.1% |
| Median latency | 14.6s | 51.5s |
Both ran with the same system prompt and the same LangChain harness. The DOCX skill was preloaded into context, so it paid no setup cost. Across both models, GPT 5.6 Terra with Vespper scored slightly higher than GPT 5.6 Sol with the DOCX skill, at about 7× lower cost. The full methodology covers the dataset, scoring, and all six solutions tested.
By task type
Where each one wins
Vespper's biggest gaps are on tracked changes, headings, and tables. The DOCX skill did better on headers and footers, though that sample is only four tasks.
| Task type (GPT 5.6 Sol pass rate) | Vespper DOCX MCP | DOCX skill |
|---|---|---|
| Tables (n=90) | 77.8% | 71.1% |
| Lists (n=96) | 71.9% | 70.8% |
| Paragraphs (n=159) | 75.5% | 72.3% |
| Headings (n=28) | 60.7% | 46.4% |
| Tracked-change management (n=26) | 57.7% | 34.6% |
| Header/Footer (n=4) | 50.0% | 75.0% |
How they work
Scripts vs a reconciler
DOCX skill
The skill gives the agent instructions for working with the .docx package. The agent writes and runs code against the file inside a sandbox. That makes it flexible, but the agent spends turns and tokens writing, running and debugging scripts. Simple things like a tracked change or a hyperlink mean working directly with OOXML.
Vespper DOCX MCP
The agent reads the document as clean HTML and sends back the HTML it wants. A model trained for this task writes the matching OOXML for each changed block. Vespper then computes tracked changes and patches the original file. The agent never touches Word XML, so its context goes to the actual task.
| Vespper DOCX MCP | DOCX skill | |
|---|---|---|
| How the agent edits | Edits clean HTML | Writes and runs scripts |
| Needs a code execution sandbox | No | Yes |
| Edits returned as Word tracked changes | Yes | If the script writes them |
| Inserting images | Not yet | Yes |
| Price | Free up to 500 edits/mo | Free, pay model tokens |
Choosing
When to use which
Stay with the DOCX skill if
- You edit a handful of documents and cost or latency doesn't matter.
- You need image insertion today.
Switch to Vespper if
- Your agent makes many edits per document, or runs live with a user waiting.
- Edits have to come back as clean Word tracked changes.
- You work on long documents with tables, lists, and inherited styles.
- You don't want to run a code sandbox just to edit Word files.
Switching
Replace the skill with one MCP connection
Connect your agent to the hosted endpoint and give it Vespper's tools instead of the skill. 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()
# Pass the tools to your agent instead of loading the DOCX skillRun your own documents through it.
The free plan includes 500 edits a month. See pricing.