Write to us

Code maps for AI

Repomap

Repomap shows an AI agent where to look in code it has never seen.

Open source Open source on GitHub

87% of the right files are in the first five picks of an AI agent new to the code 86.8% on average over 83 real tasks, in code Repomap was never tuned on. Without the agent, 69.2%. Measured on 30 September 2026

What it does

It makes two things. An overview of the whole project on a few pages: its areas, where things start and what lives where. And a task map: tell it what you want done, and it lists the files that matter, the most important first.

Nexus, our AI engineer, reads both before it touches anything.

With an agent reading it

We gave a fresh AI agent a map and asked it to name up to five files for a real task. It ran no commands and saw nothing else.

Who picks the five filesRight files found
An agent with the overview and the task map87%
An agent with the task map83%
An agent with the overview79%
Repomap’s own top five, no agent69%

Over 83 real tasks, in projects we never used to tune Repomap. Measured on 30 September 2026.

The Linux kernel

Repomap maps all 65,795 files of the Linux kernel. The overview takes 1.7 seconds and a task map 0.6 seconds, after a first run. It reads C, C++, Rust, Go, TypeScript, JavaScript, Svelte and Vue.

Where it is weakest

A task described in words the code never uses is hard for any map. There, Repomap does about as well as a plain text search.

Read the whole story