What is Universal Modder? A toolkit, not a mod manager

Universal Modder is an open-source toolkit that lets an AI coding agent mod PC games you already own. You install it into an agent such as Claude Code or Codex, point the agent at a game folder, describe the mod you want in plain language, and the agent does the rest: it fingerprints the engine, picks a route, backs up your saves, writes the code, generates the assets, tests the result in the running game, packages it, and writes down what it learned.

The project lives upstream as an MIT-licensed repository and has picked up tens of thousands of stars. This page explains what it is, what it is not, and what actually happens during a session — written for people who found the repository and want the context before installing anything.

What it actually is

Calling it a “mod” would undersell the structure. Universal Modder is three things bolted together: a set of Agent Skills, a command-line tool called um, and a shared knowledge base of field notes.

The skills are the instructions. Twelve of them ship today, led by mod-any-game, which contains the full loop and a playbook per engine family. Supporting skills cover recon, web research that survives dead forums, reverse engineering, asset generation, asset fitting, in-game automation, showcase video, mashup projects, porting mods between loaders, publishing, and writing field notes.

The um CLI is the deterministic half. Commands like um scan, um fal, um sprite, um render3d, um win, um backup and um publish check do the jobs an agent would otherwise improvise: engine fingerprinting, asset generation through the fal.ai API, sprite conversion, safe input automation, save snapshots and pre-release checks.

The knowledge base is the memory. Every finished mod can produce a field note describing the exact versions that worked, the route taken, what the engine really does, how the result was verified, and the gotchas in symptom-cause-fix form. Notes are stored upstream, so the next agent — yours or anyone’s — starts where the last one stopped instead of rediscovering the same traps.

What it is not

• Not a mod manager. There is no library view, no load-order screen, no plugin conflict resolver. If a game uses a loader such as tModLoader, Fabric or BepInEx, the agent builds a mod for that loader — but the toolkit does not manage your installed mod list the way Vortex or MO2 do.

• Not an emulator. Nothing is emulated or streamed. The agent works against the copy of the game installed on your machine, driven natively on Windows or from WSL.

• Not a cheat tool. The project explicitly refuses to inject into online games with anti-cheat, to write cheats against other players, or to bypass anti-cheat, DRM or ownership checks. Offline play and servers you host yourself are the supported scope.

• Not a one-click mod store. The output is real code, real assets and real packaging steps. Sessions can take a while, and the agent verifies in the running game rather than pretending a build log is proof.

How a session actually goes

When you ask for a mod, the agent runs the same ten-step loop every time. First it searches the knowledge base — if someone has already modded this game or engine, the plan starts from their field note. Then it runs recon to identify the engine, version, anti-cheat status, installed loaders and save locations, and turns that into a modding plan.

Before touching anything, it sets up a safe lab: save folders are backed up, and edits happen in a controlled way. It reads the game’s code — decompiling only as far as needed — then builds one working slice of the mod before expanding scope. Assets are generated with fal or converted from existing art, fitted to the engine’s exact frame size and palette rules.

The result is verified in the running game, recorded on video when useful, packaged for its platform with a publish check that blocks shipping game files or leaked keys, and finally written up as a field note for the next session. Nothing about this flow is game-specific; the playbooks change per engine, the sequence does not.

What you need to run it

An AI coding agent is the one hard requirement. Claude Code and Codex have plugin installs that bring everything in one step; Gemini CLI, Copilot, Cursor and OpenCode use extensions, marketplaces or a clone. Git, Python 3.10+ and ffmpeg need to be on PATH. uv is recommended for the CLI, Blender is only needed for 3D-to-sprite rendering, and a fal API key is optional — a local ComfyUI server through um comfy is the alternative for image generation.

The one thing it cannot do without: a game you own. Single-player titles are the bread and butter, and a local PC install is required because the agent reads and modifies the real game files on your disk.

Engine coverage

The upstream playbooks cover Unity, Unreal, .NET and XNA games (Terraria, Stardew Valley, Celeste), Godot, Source 1 and 2, Bethesda’s Creation Engine, Minecraft Java and Bedrock, Age of Empires II’s Genie engine, Mount & Blade Bannerlord, a long list of AAA native engines (RE Engine, MT Framework, FromSoftware titles, GTA, Hedgehog Engine, Black Rock, Cyberpunk 2077, Baldur’s Gate 3), plain native C++, a broad set of indie engines, and retro decompilations. Coverage is not uniform — some routes are well-trodden and verified, others are marked experimental in the field notes.

Where to go next

If you want it running today, start with the install guide and pick your agent. If you want to see the loop applied to a real game, read the first mod walkthrough. And if you are still weighing whether to try it at all, the FAQ answers the questions people actually ask before their first session.