I built a tool that makes Python coding agents 3x faster and 80% cheaper on hard refactors

Large refactors get harder as a codebase grows, and coding agents do not help the way you would hope. They act less like a senior developer and more like a resilient junior one, whose judgment runs out exactly when the refactor gets hard or the stakes get high.

The reason, I think, is that coding agents treat code like text. Files, snippets, fuzzy similarity. But code is not only text.

So I built Évariste. It precomputes a structural layer of your Python codebase and exposes it to your coding agent through MCP. The agent stops grepping naively and starts navigating your code like a senior-dev.

On large refactors, I am measuring agents running up to 3x faster and 80% cheaper in tokens, with cleaner diffs and fewer hallucinated calls.

It plugs into Cursor, Claude Code, Codex, and any MCP-compatible tool. Source code never leaves your machine. Python only for now, TypeScript and Go next.

Free for the first people on the waiting list: evariste.co. Would love feedback from anyone here who has been frustrated with their agent falling apart on a real codebase.