MCP Integration

Other AI agents

Cursor, Copilot, Codex, Gemini, and any tool without native MCP.

Two modes

For non-Claude agents, larkx supports two integration patterns:

  1. Native MCP, if the agent supports MCP (Cursor, Continue, Cline)
  2. Instruction file + CLI context, the agent reads larkx context output instead of calling tools (Copilot, ChatGPT, Codex, Gemini)

Cursor

Cursor supports MCP natively. Add larkx in Settings → MCP:

cursor mcp config
{
  "mcpServers": {
    "larkx": {
      "command": "larkx",
      "args": ["mcp"]
    }
  }
}

Then run larkx init in your project to create .cursorrules.

GitHub Copilot

Copilot doesn't expose tool-calling for arbitrary MCP servers, so it uses the instruction file pattern. larkx init creates .github/copilot-instructions.md with a directive to run larkx context at the start of any task.

OpenAI Codex

Codex reads AGENTS.md. Run larkx init and select OpenAI Codex. The generated AGENTS.md tells Codex to fetch larkx context before opening files.

Gemini CLI

Gemini reads GEMINI.md. Same flow, larkx init creates the file with the right instructions.

Check the CLI

bash
larkx --version
# → larkx v0.1.0

If the command is not found, make sure your Node global bin is on PATH.

Check the index

bash
larkx stats

Should print file counts, detected frameworks, and per-level token estimates.

If the index is empty, run larkx index first.

Check the context file

bash
cat .larkx/context.md

The file should exist and contain a compact project snapshot. If it is missing or empty, run larkx index.

Verify file-based integration

After running larkx index, confirm this file exists:

bash
cat .larkx/context.md

The file should contain a compact project map. If it is missing or empty, rerunlarkx index.

Then ask your non-MCP agent a question like:

  • "What files contain authentication logic?"
  • "Where is the login handler defined?"

The instruction file should direct the agent to read larkx context first before opening source files.

Not working?
  • If Copilot still dives into source files, confirm .github/copilot-instructions.md was created and includes larkx context.
  • If Codex or Gemini ignores the context file, reopen the terminal or chat session so it reloads project instructions.
  • For any agent, use larkx context --folder src/path or --level 1 to keep the prompt small.
  • Run larkx stats to verify the index is built and the project is recognized.

Universal fallback

For any other tool, just pipe context output into your conversation:

bash
larkx context > project-context.md
# then paste into ChatGPT / Claude.ai / wherever
Save tokens with scopingUse --folder src/auth to limit context to a subtree, or --level 1 for just file paths. Both work in any agent that can read text.