larkx indexes any project into a compact graph that Claude Code, Cursor, GitHub Copilot, and any MCP-compatible AI client can read in seconds. Measured savings on the larkx repo: 49% fewer tokens on a high-level overview query verified by larkx bench, not estimated.
larkx bench on the larkx repo (101K → 52K tokens, $0.1565 → $0.0903). Run it on your own project for your number.Four detail levels: paths, symbols, signatures, AI summaries. The AI picks the cheapest level per task.
Restrict any query to a subtree. Token cost drops roughly in proportion to the folder size, often very large savings on focused tasks.
One-line summaries per file generated by your configured provider: Local Claude or the Anthropic API. Cached locally in .larkx/summaries.json.
Works with Claude Code, Cursor, Continue, Cline, and any Model Context Protocol client.
BFS from entry points. Auto-detects Next.js routes, NestJS modules, and more.
SHA-256 file hashing. Reindex is typically much faster after the first build.
Browser explorer with Module view and Graph view. Zoom, search, VS Code deep links.
Excludes .env, *.key, *.pem, credentials* by default. Never enters AI context.
larkx init walks through AI summaries, MCP setup, agent files, and Claude Code hooks.
How many tokens does your project actually save?
larkx bench runs Claude with and without larkx and reports real token usage — not estimates.
Install, init, index, done. No accounts. No telemetry. Runs entirely on your machine.
Read the setup guideWorks with your AI tool