If you use AI coding tools, you've hit this: you start a new chat, ask Claude or Cursor to refactor something, and it hallucinates an import. Or renames a function but misses 3 call sites. Or suggests moving a file with no idea what depends on it.

The problem is simple — AI tools have no map of your codebase. Every new chat starts from zero. They burn tokens scanning files they already saw, guess at dependencies, and give you confident answers based on incomplete context. Lost context = wasted tokens + broken code + time spent fixing what AI was supposed to fix.

I built Depwire to solve this. It parses your codebase with tree-sitter, builds a complete dependency graph, and serves it to AI tools via MCP (Model Context Protocol). The graph persists across sessions — your AI never forgets the architecture.

Now when I ask "what breaks if I rename Router?", I get the exact blast radius: 5 implementing classes, 2 core framework files, 14 downstream consumers, ~25 files total. Not a guess — a deterministic answer from the actual dependency graph.

What it does: - Parses TypeScript, JavaScript, Python, and Go (tree-sitter, deterministic) - 10 MCP tools: impact analysis, dependency tracing, architecture summaries, symbol search, and more - Interactive arc diagram visualization in the browser - File watcher keeps the graph current as you edit - Zero config: npm install -g depwire-cli. No databases, no cloud, no Docker - Everything local. No data leaves your machine. - Full repository context, not just single file.

Tested on real projects — Hono (305 files, 5,636 symbols, 2.3s), Excalidraw (320 files), FastAPI, Express, Cobra. Zero parse errors.

Install: npm install -g depwire-cli GitHub: https://github.com/depwire/depwire Site: https://depwire.dev

Solo founder, scratching my own itch. BSL 1.1 (converts to Apache 2.0 in 2029). Happy to answer questions.