Vexp - Local-first context engine for AI coding agents

Hey PH, I'm Nicola, solo dev behind vexp. I was spending $400+/month on AI coding tokens and noticed something weird: most of those tokens were wasted. Every time I asked Claude Code to fix a bug, it would read 30-40 files that had nothing to do with the task. The answers were worse because of the noise, and I was paying for every irrelevant line. I looked for a solution and found that cloud-based context engines existed, but they required sending my code to their servers. That was a dealbreaker. So I built vexp. What it does: indexes your codebase into a dependency graph using tree-sitter and SQLite, entirely on your machine. When your AI agent needs context, vexp serves a ranked capsule: full source for the functions that matter, skeletonized signatures for everything connected. The agent starts every task already knowing what's relevant. I benchmarked it properly. 100 real GitHub issues from SWE-bench Verified, 4 agents, same model (Claude Opus 4.5), same budget. The setup with vexp hit 73% pass rate at $0.67/task. The most expensive setup without it: 70% at $1.98/task. 8 bugs were solved exclusively by the setup with better context. The benchmark is fully open source: github.com/Vexp-ai/vexp-swe-bench What makes vexp different from Augment Code ($270M+ funded, cloud-based): 100% local. Zero network calls. Your code never leaves your machine. Flat $19/month. No credits, no usage-based surprises. Session memory linked to the code graph. When code changes, old observations are auto-flagged stale. 30 programming languages. Dedicated tree-sitter parser for each. Works with 12 agents out of the box (Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, etc.) Free tier available, no account required. Install the VS Code extension or run npm install -g vexp-cli. Would love your feedback. What's working, what's missing, what would make you actually use this daily.