Getting Started

Index your code

Parse every file and build the graph.

Basic usage

bash

larkx index

This parses supported files, using tree-sitter for TypeScript and JavaScript when native bindings are available and a regex fallback for the other supported languages, builds the graph, and saves it to .larkx/.

Supported languages

Files with these extensions are parsed:

JavaScript / TypeScript: .js .jsx .ts .tsx .mjs .cjs
Python: .py
Go: .go
Rust: .rs
Java: .java
C/C++: .c .cpp
C#: .cs

What gets skipped

node_modules, dist, build, .next, out, coverage, .turbo, .cache
Anything in your .gitignore or .claudeignore
Security defaults from .larkx/config.json: .env*, *.key, *.pem, credentials*, etc.
Minified files (.min.js, .min.css), lock files

Incremental indexing

Each file's SHA-256 hash is stored. On subsequent runs, only files whose hash changed are reparsed. When little has changed, incremental reindexes are typically much faster than a full parse.

Force a full reindex

bash

larkx index --force

Bypasses the cache. Use this if you suspect stale data or after upgrading larkx.

AI summaries

Optional one-line file summaries are generated by the provider you configure during larkx init (Local Claude or the Anthropic API). They are stored in .larkx/summaries.json, used in level 4 context output, and shown in the visual graph UI.

bash

larkx index --ai

If you choose the Anthropic API provider, set ANTHROPIC_API_KEY first or enter it during larkx init. Local Claude uses your installed claude CLI and does not require an API key.

Summaries are cached. Once generated, subsequent --ai runs only summarize new or changed files.

Watch mode

Reindex automatically when files change:

bash

larkx index --watch

Uses chokidar. Each save triggers an incremental reparse for that file only.

After indexing

Run larkx stats to see file counts and token estimates
Run larkx serve to explore the graph in the browser
Try the AI tools, Claude Code should now use them automatically

Next: Verify setup