larkx connects to Cursor as an MCP server and indexes your project into a compact graph. Cursor can ask for just the detail it needs, file paths, symbols, signatures, or summaries, instead of repeatedly reading the same raw source files.
After running larkx init, a .mcp.json is written to your project root. Cursor reads this file on startup and connects to larkx automatically. The AI then has access to six structured query tools instead of raw file reads.
Limit any query to a subfolder. Token cost usually drops in proportion to the folder size, which helps most when you are working on one module.
Levels 1–4: paths → symbols → signatures → summaries. Cursor can choose the cheapest level that still answers the question.
SHA-256 hashing means changed files are re-indexed quickly, so the graph stays fresh without a full rebuild.
Ask which files depend on a target before changing it, which is often cheaper than re-reading the whole project.
npm install -g larkxcd my-project && larkx initlarkx indexlarkx serveDoes Cursor support MCP servers?
Yes. Cursor has native MCP support. larkx init writes the .mcp.json config that Cursor picks up automatically.
How is this different from Cursor's built-in codebase indexing?
Cursor's indexing powers its search, but the AI still reads full file content when it needs to understand code. larkx adds a structured graph so Cursor can ask for only the level of detail it needs, typically reducing repeated file reads on larger projects.
Will larkx conflict with Cursor's existing context tools?
No. larkx is additive. It exposes new MCP tools that the AI can call; it does not remove or override Cursor features.
Can I use larkx with Cursor and Claude Code at the same time?
Yes. larkx runs as a single MCP server and both editors can connect to it. The index is shared.
Is there a performance impact on Cursor?
larkx serve is a lightweight local HTTP server. CPU usage is negligible between requests. Indexing only runs when you call larkx index.
Also works with Claude Code, Continue, Cline, and GitHub Copilot
One index, every AI tool. Agents without native MCP support read a plain-text .larkx/context.md file that larkx generates alongside the graph. Setup for other agents →
Free, open source, and works best on larger or repeatedly explored codebases. Measure your own savings with larkx bench.