I got tired of the lock-in and black-box nature of AI inside MS Office. One vendor decides which models you can use, where your document data goes, and what the agent is actually allowed to do. And when it fails, you can't see the code or send a PR to add a feature you may need.

Many people don't have that choice: they're forced to use these tools (MS Office) at work. For coding agents we already have open interfaces like OpenCode where you can bring your own key, swap providers and models, run local, inspect everything, etc. I thought "agentic Office" needs the same.

So I built OpenDocBot: an agent that runs inside Word, Excel and PowerPoint. Bring your own provider (OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, Ollama, or your own endpoint / custom gateway). It's client-side, so document data goes straight to the provider you configure, and the source is public so you can always audit the code and self-host it.

I added a few things I thought would be useful: human-in-the-loop approval before any action is applied; file uploads (e.g. building a slide deck from a PDF) with local OCR for scanned documents, and tunable model behavior.

Caveat: it's early, and Microsoft Marketplace distribution isn't live yet, so today you have to sideload it. I know that's real friction, I'm fighting daily with the MS Support team to achieve MS Partner Center account validation. I wrote some scripts that automate the sideloading process, and recorded short videos to make it easier.

If you use Office heavily, I'd love your feedback.

Repo: https://github.com/opendocbot/opendocbot

Site + docs: https://opendocbot.com