repowise

Repowise gives an AI coding agent a real map of your codebase. It builds five layers of intelligence (a dependency graph, git history analytics, auto-generated docs, architectural decisions, and a per-file code-health score) and serves them to Claude Code, Codex, or any MCP client through nine focused tools. Instead of your agent grepping around blindly, it can ask what is risky here or why this code exists and get a grounded answer. AGPL-3.0, free for internal use.

It runs locally: `pip install repowise`, then `repowise init` to index a repo using Tree-sitter across 15 languages. There is no cloud service; it lives on your machine or in CI and exposes an MCP server. Ops is light for a single repo and more involved if you index many large ones or wire it into a whole team's agent setup. The health scoring is deterministic, 25 markers per file rather than an LLM guessing, which is the genuinely interesting part.

For anyone running AI agents against a nontrivial codebase, this is a real upgrade to how much the agent actually understands, and it is free for internal use. Solo developers and small teams use it as-is. Larger companies that cannot take on AGPL obligations can buy a commercial license to escape the copyleft terms, which is the only place money enters.

The catch is that the impressive numbers (defect prediction accuracy, "2.3x more defects than CodeScene," big token savings) are the project's own, measured on its own test set. They may hold up, but treat them as claims to verify against your repo, not settled fact. The tool is useful for agent context either way; just do not buy the benchmark line without testing it on your own code.