Frequently asked questions

larkx is a CLI tool that parses your codebase with tree-sitter, builds a queryable graph of every file, function, class, import, and call relationship, then makes it available to AI agents in three ways: as an MCP server (6 query tools for MCP clients), as a pre-generated .larkx/context.md file (readable by any agent), or via the larkx context CLI command.

Approximately 60-85% on typical projects. Example: a 200-file project costs ~16,000 tokens via larkx level 2 versus ~120,000 tokens reading every file directly. Searching for one function costs ~30 tokens with search_symbol versus ~600 tokens reading the file. Actual savings vary by project size, task, and AI model.

A pre-generated plain-text codebase map written automatically on every larkx index run. Any AI agent that can read files uses it — no MCP server needed. It contains the same level-2 output as get_project_index: file paths, function names, line numbers, and imports at ~80 tokens per file. It is gitignored.

MCP clients (Claude Code, Cursor, Continue, Zed, Windsurf) use the 6 MCP tools. All other agents (GitHub Copilot, OpenAI Codex, Gemini CLI) read .larkx/context.md directly. Any agent or tool can also use larkx context CLI output by pasting it into the chat.

No. The indexer, MCP server, context.md generation, and CLI all work with zero API keys. AI summaries (larkx index --ai) are optional and require either the claude CLI (free with your Claude subscription) or an Anthropic API key.

AI summaries generate a one-sentence description per file, stored in .larkx/summaries.json. Choose Local Claude (uses the claude CLI with your Claude Pro/Team subscription, no extra API key) or Anthropic API key (billed per token, uses claude-haiku-4-5). Skip is the recommended default — summaries cost extra tokens to generate.

TypeScript, JavaScript (.ts .tsx .js .jsx .mjs .cjs), Python, Go, Rust, Java, C, C++, C#, Ruby, Swift, Kotlin, Scala, and Shell. Tree-sitter is used where native bindings are available, with a regex fallback for everything else.

Yes. MIT licensed. Free for personal and commercial use. No telemetry, no SaaS, runs entirely on your machine.

No. Everything stays local. The only optional external call is larkx index --ai, which sends each file to the LLM provider you configured (Local Claude via claude CLI, or Anthropic API via your own key).

Auto-detection. If next is in package.json, larkx treats app/**/page.tsx, app/**/layout.tsx, route handlers, and middleware.ts as entry points for reachability analysis. They never appear as dead code.

No. The default exclude list includes .env, .env.*, *.key, *.pem, *.cert, credentials*, secrets/, and *.token. These files are never parsed or indexed.

A 1,000-file TypeScript project takes 10-30 seconds on the first index. Subsequent runs are incremental (SHA-256 hashed) and typically under 2 seconds. The context.md file is regenerated on every run.

Make sure .mcp.json exists in your project root, then Ctrl+Shift+P → Developer: Reload Window. Claude Code will prompt to approve the server — click Allow. Run larkx mcp --check to verify the server responds.

A Claude Code hook written to .claude/settings.json that runs before every prompt. With MCP enabled it reminds Claude to use larkx MCP tools first. Without MCP it tells Claude to read .larkx/context.md first. larkx init sets this up automatically.

Yes. Run larkx init and larkx index from the monorepo root. Use --folder to scope get_project_index to a specific package when working in one area.

Yes. Fully supported on Windows 10/11. Path separators are normalized automatically.

Not yet. .vue and .svelte parsers are on the roadmap. Those files are skipped for now.