Previously, I had Claude and Cursor handle Word, Excel, and PPT mainly with:

Python libraries + XML + templates + a pile of prompts.
It works, but it’s especially easy to mess up.

Today I came across a GitHub project with 13 thousand stars, and the approach is completely different.
OfficeCLI directly turns Word, Excel, and PPT into a suite of AI Agent-callable CLIs.

The best part isn’t editing Office—it’s that it finally makes Office visible to AI.
Previously, when Agents changed PPT, they basically relied on guessing; after editing, you didn’t know what it looked like, so you had to keep guessing.

OfficeCLI includes an HTML rendering engine, so it can render Word, Excel, and PPT directly into HTML or PNG.
So the whole workflow becomes:
Edit → Render → Check → Fix.
This is a real Agent workflow.

A few other design choices are also really comfortable:
1️⃣ Path-based operations
Unlike before, where you spent forever hunting nodes in XML.
Now you can directly:
locate elements, and change exactly what you want.
The AI basically won’t get lost.

2️⃣ Built specifically for Agents
All commands support JSON, and even errors are JSON.
If the path is wrong or the parameters are incorrect, it tells the Agent where the mistake is, and also provides correction suggestions.
The Agent can then continue to the next step by itself.

3️⃣ MCP out of the box
Claude Code and Cursor can register the tool directly.
After installation, it can also automatically set up Skills, so you basically don’t have to tinker.

Excel isn’t just about filling cells either.
It supports 350+ functions, formulas, charts, pivot tables, conditional formatting, and data validation.
Word and PPT also have full capabilities—not just changing text.

I think its real value isn’t Office automation itself.
It’s that for the first time, someone has made Office into a tool that an Agent can call reliably, and that can also verify the results.

In the future, writing weekly reports, creating PPTs, and managing financial sheets will very likely look like:
One sentence ➜ and the Agent calls OfficeCLI by itself ➜ to check the result, then ➜ fixes it until satisfied.
That’s what an AI Agent should feel like.
🔗 GitHub: