圭司

midown

i tried out midown, a tool that helps AI coding agents understand how a codebase is put together.

it also happens to be built by a friend from college. we worked on a road defect indexing system together, it's fun to see something he made grow into its own project.

what is midown?

midown turns your codebase into a graph.

instead of letting an AI dig through every file to figure out how things connect, midown hands it a clear view of your files and how they relate.

it can also share that info over MCP, so tools like claude and cursor can read it.

midown basically gives your AI coding agent a map of your code.

midown's graph of a codebase, files and components joined by lines
the graph view: files and components, joined by what they use

getting started

install the CLI:

npm i -g @bentlyr/midown

go to your project:

cd my-project

add it to midown:

midown add .

then index it:

midown push

you should see midown go through your files and build the graph.

a terminal running midown add and midown push
midown add . then midown push, indexing the files

connect your AI tools

once the project is indexed, run:

midown start

that starts the local MCP server, which lets supported AI coding tools read your codebase structure.

a terminal running midown start
midown start, bringing up the local MCP server

see your project

you can also run:

midown visualize

this opens a local web app where you can click around and see how files and components connect.

the midown visualizer open in a browser
the local visualizer, with one node selected to show what it depends on

why bother?

the point is less searching and less context for the AI.

on a small project you probably will not notice much. on a big one, a map helps the AI find the right files faster.