Breaking

Repowise: open-source tool indexes codebases into five layers

July 4, 2026 at 10:22 EDT

Software Dev & Coding
AI Agents
Open Source

Developer Tom Dörr and collaborators have released Repowise, an open-source tool that indexes codebases into five "intelligence layers." It feeds structured context to AI agents such as Claude Code and Cursor via MCP, while also producing code health scores and refactoring plans for human users.

Open Source · Developer Tooling

Repowise: five intelligence layers that give AI coding agents real codebase context

An open-source tool (AGPL-3.0) that indexes a repo into SQLite + a vector store in one pass, then serves structured context to agents like Claude Code through nine task-oriented MCP tools.

intelligence layers indexed in a single pass

task-oriented tools exposed over MCP

<30s

to index a 3,000-file repo, no LLM required

Defect detection under a fixed review budget

Same time to review — how much more gets caught.

1×

Baseline
review

2.3×

With Code Health
layer

Popt +0.144 · recall +0.098

Defect prediction (ROC AUC)

0.74

Repowise

0.705

CodeScene

Across 21 repositories, 9 languages.

SWE-QA efficiency

-96%

tokens used

-70%

tool calls

With comparable answer quality.

The five intelligence layers

Graph

Dependency graph across 15 languages via tree-sitter; module partitioning with Leiden community detection.

Git

Risk indicators — hotspots, ownership, co-change and bus factor.

Docs

Incrementally updated LLM-generated wiki, searched via hybrid RAG.

Decisions

Mines design decisions from 8 sources and tracks staleness.

Code Health

Scores 1–10 using 25 deterministic markers for defect risk, maintainability and performance.

Developers report

Better agent context, fewer file reads
Useful decision tracking & dead-code detection
Shorter onboarding, more concrete refactor plans

Open questions

Why cap MCP tools at 9 when 30+ are possible?
Wiki relies on freshness & confidence scoring
Privacy in hosted use (BYOK supported)

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.