I built this because AI coding assistants waste enormous amounts of context on code navigation. Every grep returns hundreds of
text matches that the AI reads through — consuming 2000+ tokens for a simple "Where is function X?" question.
AiDex pre-indexes your codebase with Tree-sitter and serves results via MCP (Model Context Protocol). A search returns ~50
tokens instead of 2000+. It knows the difference between identifiers and text — searching for log won't match catalog.
Technical stack: Tree-sitter for parsing, SQLite (WAL mode) for storage, MCP stdio transport. Supports 11 languages. Index time
is ~1s per 1000 files, queries take 1-5ms.
It's not a vector DB or embedding-based search — it's a plain identifier index with exact matches, prefix/substring search, and
method signature extraction. Simple, fast, deterministic.
Also includes a built-in task backlog (persisted in SQLite) and an interactive browser-based viewer for exploring the index.
Works with Claude Code, Cursor, Windsurf, Gemini CLI, VS Code Copilot, and anything else that speaks MCP.
Update: v1.9.0 – Cross-platform screenshots
Just shipped a new feature: AiDex can now take screenshots directly via MCP. Four modes: fullscreen, active window, specific window (by
title), and interactive region selection.
The killer use case: error dialogs or UI bugs where you can't even copy the text. Your AI assistant captures the region, reads the image,
and immediately understands the problem. No manual file paths, no clipboard gymnastics.
Works on Windows (PowerShell + .NET), macOS (screencapture), and Linux (maim/scrot). No new dependencies.
aidex_screenshot({ mode: "window", window_title: "VS Code" })
aidex_screenshot({ mode: "region" }) // user draws a rectangle
aidex_windows({ filter: "chrome" }) // find window titles
npm update -g aidex-mcp to get it.
2 comments
[ 3.1 ms ] story [ 16.6 ms ] thread