Your AI agent needs a map, not just a file list

What it does repowise scans a repository and keeps a live-updated index of five layers: a tree-sitter dependency graph across 15 languages, git-derived hotspots and ownership, LLM-generated docs rebuilt per commit, architectural decisions mined from eight sources, and a deterministic code-health score. It exposes all of this to MCP-compatible agents through nine task-specific tools, and includes a built-in web dashboard for human browsing. The goal is to replace repetitive agent exploration—re-grepping, re-reading files—with pre-computed context.

The interesting bit The code-health layer is the standout: it scores every file 1–10 from 25 deterministic biomarkers (complexity, churn, clone detection, etc.) calibrated against a real defect corpus, with zero LLM calls and no cloud dependency. In a head-to-head on 2,770 files, it surfaced 2.3× more defects under a fixed review budget than a leading commercial tool, and the README notes this is the only layer they can prove predicts actual bugs.

Key highlights

• Agent efficiency: Benchmarks on public repos claim −49% to −70% tool calls and −69% to −89% file reads per query versus raw agent exploration, with answer quality at parity.
• Deterministic health scoring: Runs in under 30 seconds on a 3,000-file repo using pure Python and tree-sitter; weights are learned from a defect corpus, not hand-tuned.
• Architectural decisions: Mines decisions from eight sources, links them to code nodes, and tracks relationships like supersedes and conflicts_with—a layer the authors say is captured nowhere else.
• Local-first: Runs fully offline, serves a local web UI, and supports multi-repo workspaces.
• Language coverage: Full-tier parsing for nine languages (Python, TypeScript, Java, Go, Rust, etc.) and good-tier for six more.

Caveats

• The efficiency benchmarks show a wide range (−49% to −70% tool calls, −29% to −36% cost), and the “best case” −70% / −89% figures are the top of that range.
• Only nine of the 15 supported languages are at “Full tier”; the remainder are “Good tier,” which implies uneven depth.
• The health-score comparison against a commercial competitor is on a specific 2,770-file cross-language set; your mileage may vary on smaller or monolingual codebases.

Verdict Worth a look if you run AI agents against large, unfamiliar repos and want to cut token burn and repetitive exploration. Probably overkill if you already know every file by heart or your codebase fits in a single prompt.