MCP Server

Overview

codegen ships with a Python MCP (Model Context Protocol) server, codexx-mcp-server, that exposes the C++ AST and the codegen rule system to LLM agents. This is the substrate for LLM-driven code intelligence: an agent can query the AST of any header, inspect entity metadata, and invoke generation rules, all via standard MCP tool calls.

What it enables

• AST introspection: An LLM agent can ask “what are the public fields of ConnectionOptions?” and receive structured JSON, not text parsed from a header.
• Rule invocation: An agent can trigger a codegen rule run and receive the generated output for review before it is written.
• Schema-guided generation: The JSON AST schema is deterministic and documented. An agent can reliably generate new codegen rules from examples without guessing at the data shape.

Starting the server

codexx-mcp-server --input ./include --port 3000

Or via stdio transport for direct LLM client integration:

codexx-mcp-server --input ./include --transport stdio

Available tools

Tool	Description
get_entity_schema	Return the JSON schema for a given node kind (Struct, Enum, etc.)
run_dtdk_command	Invoke a codegen rule and return the generated output

Claude Code integration

Add to your mcp_servers config:

{
  "mcpServers": {
    "codegen": {
      "command": "codexx-mcp-server",
      "args": ["--input", "./include", "--transport", "stdio"]
    }
  }
}

Claude Code can then call get_entity_schema to understand your codebase structure and run_dtdk_command to preview generated output, making the rule-writing workflow interactive.

Key Takeaways

• The MCP server exposes the codegen AST and rule system to LLM agents via standard MCP tool calls.
• JSON AST schema is deterministic, LLMs can reliably generate new rules from examples.
• get_entity_schema returns the schema for a node kind; run_dtdk_command invokes a rule.
• Use --transport stdio for direct integration with Claude Code and other MCP-compatible clients.