Six months ago I was deep into a feature session when I noticed something: I was spending more prompts explaining my codebase to the AI than actually building.
"No, that function is used in three other places."
"No, we did it that way on purpose because of X."
"Check the other file first."
The same corrections, session after session.
As your project grows past that early magic phase, the AI has no map. It doesn't know what's connected to what or why things are shaped the way they are. Context overhead crowds out progress.
Repowise is my fix for that.
You might also like
Why does AI context become a bottleneck as your project grows?
The model reads files one at a time with grep. It doesn't know what's connected to what, and it doesn't know the history of why things are the way they are.
Every session you end up re-explaining architecture decisions the AI should already know.
What is Repowise and how does it give AI a map of your codebase?
I built an open source tool called Repowise that gives the AI that map. It's an MCP layer — works with Claude Code, Cursor, Codex, anything MCP-compatible.
You run pip install repowise && repowise init and it indexes your entire project into five layers the model can query before it starts writing code.
What are the five layers Repowise uses to understand your project?
- Graph — Knows what depends on what. Before the AI touches a file, it already knows what's connected.
- Git — Knows which files always change together, even when there's no import between them.
- Docs — Auto-generated, searchable wiki of your whole codebase.
- Decisions — Knows why your code is shaped the way it is. That weird workaround you asked the AI to write last week? It won't "fix" it next week.
- Code Health — 15 Biomarkers that score every file 1–10, flagging the messy ones before the AI makes them worse.
Repowise also auto-generates and keeps your .claude.md (or equivalent rules files) updated, so the AI starts every session already knowing your project structure.
What were the results of the Django experiment?
We ran a time-travel experiment on Django across 542 files: scored every file with the Code Health layer, then counted bug-fix commits over the next six months.
14 of the 20 worst-scoring files had real bugs. The top predictors were untested hotspots and developer congestion — not complexity metrics.
How does Repowise improve AI code generation in benchmarks?
My benchmarks show:
- 49% fewer tool calls
- 89% fewer file reads
- 36% cost reduction
There's also a local web UI where you can see your whole project visually — dependencies, hotspots, dead code, all of it.
2K+ stars. Open source. Self-hosted. Free.
https://github.com/repowise-dev/repowise
FAQ
What languages does Repowise support?
The pip install repowise entry point is Python-native, and my benchmarks used a Django project. The MCP layer itself is compatible with any MCP-capable editor — Claude Code, Cursor, Codex — so it can index whatever codebase those tools point it at. I've tested it most thoroughly on Python projects; check the GitHub repo for current language coverage.
How is the Code Health score calculated?
Each file gets a score from 1 to 10 based on 15 Biomarkers. In the Django experiment, the top predictors of real bugs were untested hotspots and developer congestion — not raw complexity metrics. The score surfaces which files are risky before the AI touches them, so you can triage proactively rather than after a regression.
Is Repowise free for commercial use?
Yes. Repowise is open source and self-hosted. There's no SaaS tier or usage fee — you run it on your own infrastructure. The GitHub repo has over 2K stars and is the authoritative source for the license terms.
0 Comments
Log in to comment
Not a member yet? Join the community
Pick a meme
KlipyHave a great take?
Drop your email — we'll send a magic link so you can post it. No password.
Not a member of the community? Join today.
Join the community →