mcp-ue/CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Purpose

UnrealDocGenerator is a tool for generating documentation from Unreal Engine C++ header files and exposing it to Claude via an MCP server for item-granularity lookups.

Current State

Implementation complete. Scripts live in docgen/:

• docgen/ue_parser.py — Parses UE headers into dataclasses
• docgen/ue_markdown.py — Renders parsed data as Markdown (ultra-compact format, documented items only)
• docgen/generate.py — CLI entry point; two-pass pipeline (parse-all → build type index → render-all)
• ue_mcp_server.py — MCP server exposing 5 tools for item-granularity doc lookups
• .mcp.json — Registers the MCP server with Claude Code (stdio transport)

Usage

python docgen/generate.py <input> [input2 ...] <output_dir>

python docgen/generate.py Runtime/Engine/ Runtime/AIModule/ docs/   # multiple directories
python docgen/generate.py Runtime/Engine/Classes/GameFramework/Actor.h docs/  # single file

Output: one .md per .h + docs/type-index.txt (compact TypeName: path/to/File.md lookup).

The last argument is always the output directory. All preceding arguments are inputs (files or directories, processed recursively).

MCP Server

ue_mcp_server.py exposes 5 tools Claude can call directly:

Tool	Purpose
search_types(pattern)	Regex search over type-index.txt — locate a type before fetching details
get_class_overview(class_name)	Compact view: description + inherits + property/function name lists
get_member(class_name, member_name)	Full doc for one function or property (all overloads)
get_file(relative_path)	Full .md file — use when you need delegates, enums, or multiple classes
search_source(pattern, path_hint)	Grep UE .h files in Engine/Source/ and Engine/Plugins/; path_hint applied under both; returns up to 40 lines total

Requires pip install mcp. Reads $UE_DOCS_PATH; search_source also requires $UE_ENGINE_ROOT (must point to the Engine/ directory, e.g. /path/to/UnrealEngine/Engine).

.mcp.json passes both vars via ${VAR} interpolation — no paths hardcoded.

Typical query flow: search_types → get_class_overview → get_member (~3 calls, ~15 lines vs ~100 lines for a full file read).

Architecture Notes

• Parser uses a position-based scanner, not line-by-line regex, to handle nested braces correctly.
• find_matching_close() extracts balanced {}/() while skipping //, /* */, and string literals.
• Only items with actual C++ doc comments (/** */) are included in output (_has_doc() filter).
• Two-pass pipeline in generate.py: Pass 1 parses all files, Pass 2 renders with cross-reference links.
• Multi-input: each (header, base) pair tracks its own base dir for correct relative output paths.
• Inline cross-references: base class names linked to their .md when in corpus (_make_type_link()).
• Deprecated functions excluded from output (not f.is_deprecated in visibility filter).
• Enums with no value descriptions use compact inline format (Values: A, B, C) instead of a table.
• Placeholder descriptions (------//) filtered by _PLACEHOLDER_RE in _fn_body().
• Headers with no documented content: render_header() returns "", generate.py skips writing the file and filters those types from type-index.txt to avoid dead entries.

Critical Gotchas

• Never use \s inside a [...] character class with *? on C++ source — causes catastrophic regex backtracking. Use line-based scanning instead.
• Identifier regex must be r'\w+', not r'[\w:~]+' — the latter consumes public: as one token, breaking access specifier detection (all members become private).
• Module name casing: use _caps_to_camel() (not .title()) — AIMODULE.title() → Aimodule instead of AIModule.