• 04 Nov, 2025 *

The company wiki is ubiquitous.

And when someone says, the information you need is in our wiki. This is what I think:

Because the information I need isn’t going to be there and if it is it’ll be 6 months out of date.

This might be solvable with:

A) hyperlinks B) vector search across the pages C) Super diligence, where everyone remembers to update the document every time

But in practice none of these work. 1

But how close does vector search get?

It understand semantics. And with MCP it’s possible to do flexible search over text.

If the goal is to create an up to date ‘wiki’ page in ~5 seconds, then it seems like a good MCP server with the right data in a vector store could do the job.

So, I built a tool called flow to do this.

It’s here. And it’s free and open source.

Flow is a tool that uses OCR to log your work and then allows you to use it to with your favorite MCP frontend later. This is done with a set of useful tools built on top of ChromaDB that has JSON blobs of your screen history.

In short, it’s an MCP server.

And it all runs locally so the data is yours until you plug it into your client side MCP.

Then it’ll get blasted out to OpenAI, Anthropic, etc.

Note: you’ll need to open a few terminal tabs to run all the processes required and it’s only tested on Mac.

Also Note: This is alpha software so there might be breaking changes at any times.

Recently I’ve added localhost website and ngrok support so that you can:

1. Start flow
2. Open your MCP client
3. Ask for an answer to a question about something you worked on in the past
4. Ask it to create a page - these pages are thin wrappers around markdown files that are stored on disk
5. Share via your custom ngrok endpoint

Thanks for reading.

A) only works if everyone knows where all the pages are already - so in practice no one maintains good links. B) falls apart the moment there are multiple systems at play. C) doesn’t work for the obvious reasons.↩