universal-modder is a skills, tooling and knowledge-base package that lets AI coding agents like Claude Code and Codex mod PC games their owners control, wrapping recon, reverse engineering and asset generation into one documented loop.
| Tool | rehan-remade/universal-modder — skills, an fal MCP server and the um CLI that let AI coding agents mod PC games |
| Category | AI agent tooling / game reverse engineering framework (Python, MIT) |
| Primary Use | Automating game modding on games you own: engine recon via um scan, code reading, asset generation with um fal, in-game verification, and shared field notes in knowledge/ |
| Safe Use | Single-player or otherwise authorized modding of games the operator owns, in lab setups with backed-up saves; CONTRIBUTING.md explicitly forbids cheating other players |
| Telemetry Note | Purely local against game installs; the only outbound calls are fal.ai generation requests under FAL_KEY (or a local ComfyUI server via um comfy), so defenders observe normal agent traffic plus those API calls |
universal-modder sits at an interesting intersection: it is nominally a game-modding framework, but architecturally it is a full agentic workflow harness — skills, an MCP server, a Python CLI, and a versioned knowledge base that agents themselves write to via pull requests. The premise is that any capable coding agent (Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode) can be pointed at a PC game you own and walked through the entire mod lifecycle: identify the engine, pick a route, read the actual code, build one working slice, generate art and audio, verify in the running game, record a showcase video, package the mod, and leave a field note behind. For security professionals, the interesting part is less the mods and more the disciplined recon-and-reverse-engineering loop the project encodes, which is a clean reference implementation of how to structure agent-driven analysis of an unfamiliar binary target.
The repository has real traction — roughly 3,900 stars, MIT licensed, Python as the primary language, with topics spanning claude-code-plugin, mcp, modding, reverse-engineering and specific games like tmodloader and age-of-empires. The README is unusually detailed, laying out install paths per agent, a ten-step operating loop, a skills catalog, the um CLI surface, and the contribution contract. This is a project designed for humans to supervise agents, not for fire-and-forget automation, and that framing shows throughout.
Installation is deliberately multi-agent. Claude Code users install via /plugin marketplace add rehan-remade/universal-modder and /plugin install universal-modder@universal-modder; the generic fallback is simply git clone https://github.com/rehan-remade/universal-modder and starting your agent inside it, since instructions live in AGENTS.md (pointed to by CLAUDE.md and GEMINI.md). The skills are duplicated into .agents/skills and .claude/skills so each agent finds them where it expects, and MCP configuration is scattered deliberately across .mcp.json, .codex/config.toml, .cursor/mcp.json, .vscode/mcp.json and opencode.json. The um CLI can be installed standalone with uv tool install git+https://github.com/rehan-remade/universal-modder.
The core loop, driven by the mod-any-game skill, is worth studying as agent architecture. It always runs the same ten phases: knowledge-base search, recon and route selection, safe-lab setup with saves backed up, reading real code, building a minimal working slice, asset generation, verification in the actual game, recording, packaging, and writing a field note for the next agent. That last step is the project's most original idea — knowledge/ holds notes that record the exact versions that worked, the chosen route and why, what the engine really does, how it was verified, and gotchas in symptom-cause-fix format. Agents that finish a mod can open a PR with um kb pr, so institutional knowledge accumulates instead of evaporating between sessions.
From a reverse-engineering perspective, the game-recon and reverse-engineering skills are the meat. um scan locates Steam/Epic/Xbox installs, fingerprints the engine and version, distinguishes managed (.NET) from native code, detects anti-cheat, inventories installed loaders and save folders, and ranks candidate modding routes into a MODDING_PLAN.md. The reverse-engineering skill then wires the agent into ILSpy, Cpp2IL, Vineflower, Ghidra and IDA over MCP, plus Cheat Engine, Frida and RenderDoc, and requires proving any reverse-engineered file format with a round trip. That round-trip proof requirement is good engineering hygiene that many human RE workflows skip.
Asset generation is handled through fal, exposed both as an MCP server and as the um fal command with subcommands like sprite, pixelate, texture, pbr, model3d, rig, sfx, music, voice and video, all plain REST with a manifest of every generation. If you do not want an API key, um comfy routes image generation through a local ComfyUI server instead — a thoughtful offline alternative. The asset-pipeline skill converts that generated art into engine-exact frames: cutouts, nearest-neighbour fits, palettes, sprite sheets, team-colour masks, and 3D-to-8/16-heading sprite renders via Blender. External dependencies are minimal and sensible: Git, Python 3.10+, ffmpeg, and optionally Blender.
The engine coverage is broad: the README claims twelve playbooks inside mod-any-game covering Unity, Unreal, .NET/XNA titles like Terraria and Stardew Valley, Godot, Source 1/2, Bethesda games, Minecraft, AoE2/Genie, RE Engine/FromSoft/GTA/Cyberpunk/BG3-class native C++ targets, indie engines (GameMaker, RPG Maker, Ren'Py, Paradox, Doom, HTML5, LÖVE, Java), and retro decompilation projects. The mashup-mods skill extends this to game-inside-game scenarios, with community examples including real Minecraft inside Portal 2 and GTA V story mode. The showcase examples read as ambitious, but each is backed by a knowledge-base note documenting how it was actually achieved and verified.
Safety posture is explicitly addressed, which matters for anything touching game memory and anti-cheat. The game-automation skill emphasizes GPU-safe screenshots, safe clicking/typing, windowed mode and crash-reporter cleanup, and step three of the loop mandates a safe lab with backed-up saves. More importantly, CONTRIBUTING.md codifies the rules: no game files, no decompiled dumps, and no cheating other players — effectively confining the tool to single-player or otherwise authorized modding. The project also forbids uploading copyrighted content to the knowledge base, keeping it as methodology rather than payload. That is the right line, and it is the same boundary professionals should respect: mod your own games, offline or on servers where modding is sanctioned.
For defenders, the telemetry footprint is worth understanding. Everything um does against a game is local — file scans, decompilation, memory inspection via Cheat Engine or Frida — and the only significant outbound traffic is generative-asset calls to fal.ai under FAL_KEY, or none at all if um comfy is used. Anti-cheat vendors will still see the same injection and debugging primitives as any modding workflow; the novelty here is orchestration, not primitives. Security teams evaluating agent platforms may find the project useful as a case study in scoping agent capabilities: the MCP surface is narrow, the CLI is inspectable Python, and every generation is manifest-logged.
Criticisms are minor. The project leans heavily on the fal commercial API for its showcase asset quality, and the um CLI is the glue that could break if upstream agents change their plugin formats — though the plain git clone fallback is robust against that. The knowledge-base model, where agents open PRs after a human says OK, correctly keeps a human in the publication loop. As a piece of agentic engineering and a documented, ethical reverse-engineering workflow for software you own, universal-modder is one of the more instructive repositories of its kind, and well worth a read even if you never mod a game.
rehan-remade/universal-modder.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.