Co-op Anything is a GitHub repo that does not contain a game mod. It contains instructions for your AI agent. The idea in its README is blunt: point Claude Code, Codex, Gemini CLI or Cursor at the repo, say which single-player game you want to play with friends, and the agent researches what exists, reverse engineers the game if it must, builds or fixes the co-op, tests it with bots inside the game, and packages it for one-click install.
The repo, cranebeaker/coop-anything, was created on October 16, 2026. By October 30 it had about 1,598 stars, 222 forks, 410 commits and an MIT license, per the GitHub page and API. It is a good example of a growing pattern: a repository that is mostly a playbook for agents rather than code for humans. Below is what is in it, what has actually shipped, and the caveats.
What Co-op Anything is, in one paragraph #
It is three things bundled together. First, a playbook of Markdown guides on topics like how co-op mods work, engine families, reviving abandoned mods, networking and desync hunting. Second, a set of agent skills (in .agents/skills, mirrored in .claude/skills for Claude Code) such as coop-research, coop-revive, coop-reverse-engineer, coop-build-from-zero, coop-networking, coop-sync-testing, coop-qa and couchlink. Third, an MCP server and a toolbox of sync-testing utilities. The repo's AGENTS.md is the entry point that tells an agent which phase it is in.
If you are new to the terms, our explainer on agent skills as an open standard covers how a skill file steers an agent, and the REA setup guide covers the reverse-engineering side that Co-op Anything leans on.
The quick start #
According to the README, you tell your agent:
Read https://github.com/cranebeaker/coop-anything and follow AGENTS.md. I want to play (game) in co-op with my friends.
The agent then follows the skills. To give it structured tools, you add the Co-op Anything MCP server. Claude Code picks it up from the repo's .mcp.json; other agents use claude mcp add coop-anything -- python <clone>/tools/mcp/coop_mcp.py, with Codex, Gemini CLI and Cursor setup in playbook/mcp.md. The README says the MCP lets an agent search the co-op directory, detect a game's engine from its install folder, read the playbook and skills, and validate contributions.
For the binary-level work, the README recommends pairing it with REA, the reverse-engineering MCP (npx rea-agents setup): "Co-op Anything knows what to do, REA looks inside the game." We covered REA's security model in our REA MCP explainer, and Co-op Anything was already one of the entries in our top 10 projects built with REA list when it had about 437 stars. It has roughly tripled in attention since.
Two paths: finish what exists, or build from zero #
The README splits the work into two paths.
| Path 1: finish what exists | Path 2: build from zero | |
|---|---|---|
| When | A co-op mod exists but is buggy, stalled or abandoned | Nothing usable exists |
| How | Find the last build, mine bug reports (issues, forums, Discord), audit the code, fix it as a separate package, test, ship | Reverse engineer the game and engine, hook in, show the second player, sync enemies, world and story, handle level changes and saves, test, ship |
| Skills | coop-research ,coop-revive ,coop-qa ,couchlink | coop-reverse-engineer ,coop-build-from-zero ,coop-networking ,coop-sync-testing ,coop-qa ,couchlink |
The README claims that when a mod already exists, "most projects take days at most", while from-zero depends on the engine. That is the maintainer's estimate, not a measured benchmark; treat it as optimistic.
Approaches by engine #
The most useful part for a curious reader is the table of typical approaches by engine family. From the README:
- Engines with multiplayer already (Unreal 1/2, Source, id Tech, X-Ray): run the single-player logic on the existing netcode with script mods.
- Bethesda (Gamebryo, Creation): a script-extender plugin with host-authority sync, using tools like SKSE, F4SE, xNVSE, Ghidra and REA.
- Unity: Harmony patches through a mod plus a C# networking library (BepInEx, Melon, dnSpyEx, Il2CppDumper).
- Unreal Engine 4/5: Lua or C++ mods with an SDK dump of the game's classes (UE4SS, Dumper-7).
- Decompiled or open-source games (N64 decomps, KeeperFX, OpenMW): add networking directly in the code.
- Console-only titles: a decomp or recompiled PC port first, otherwise emulator netplay or memory sync between two emulators.
- Closed binary, nothing else: full reverse engineering of state, saves and levels, which the README itself calls the hardest path.
This lines up with the decompilation trend we have covered in AI-decompiled game clones running in browsers and the Snowboard Kids 2 workflow breakdown, where the same pattern holds: agents are strongest when there is already a decomp or mod ecosystem to build on.
The sync-testing toolbox #
Making a game multiplayer is mostly a testing problem: two copies of the world drift apart, and the bug appears only on one machine. Co-op Anything ships game-agnostic tools in tools/, each with tests, "mutation twins" and a single command ( python -I tools/test_all.py) that prints ALL OK. They come from the maintainer's fnv-coop project. Highlights from the README:
- chaos: a UDP relay that adds lag, loss, reordering, duplicates and one-way cuts between real clients.
- flight: a flight recorder plus analyzer that lines up two to four machines, finds the first desync, shrinks it and writes a regression test.
- fuzz: a seeded N-player fuzzer with invariants such as "converge", "no double apply" and "one owner".
- ownership: a small C++ library for one owner per NPC with epochs, handoff and stale-echo rejection.
- animcheck: compares animation and pose across machines and flags sliding, T-poses and rubber-banding.
- hotreload: swaps a mod's core inside running games instead of restarting them.
This is the piece that makes the "agent does the work" claim more plausible than a bare prompt would. An agent that can run a chaos relay and a desync analyzer has a feedback loop. For a broader look at why verification matters more than generation here, see our piece on agents with hard guardrails and checks.
What has actually shipped #
The README's "Games so far" table lists five release repos, all under the maintainer's account:
| Game | What was done |
|---|---|
| Deus Ex (2000) | Revived the HX co-op mod: fix package, Steam relay, automated test suite |
| Dungeon Keeper (1997) | Shared-keeper co-op on KeeperFX, Steam relay networking |
| S.T.A.L.K.E.R. Shadow of Chernobyl | Fix patch for Shadow of COOP, with the true ending unlocked |
| Fallout: New Vegas | Story co-op: shared quests, dialogue results, karma and reputation |
| Zelda: Ocarina of Time | Fixes for a fully synced Ship of Harkinian co-op fork |
It also lists community entries (Mega Man X and X4, Pilotwings 64 split-screen, Crab Champions, Stacklands, Half Sword, Enter the Gungeon, Divinity II, Command and Conquer 3, Far Cry 2, Gothic 3, Modern Warfare 2 campaign and others) and a directory of 230+ co-op mods and gaps, browsable at coopanything.app, with a QA level per entry: untested, player reports, automated tests with bots driving real clients, or verified playthrough.
That QA label is the honest part of the project. A listing does not mean a polished mod. We did not install or play these builds, so we cannot vouch for their quality; check the QA level and the release repo's issues before you invite friends.
## Couchlink: same-PC split screen
If your friends are on the couch rather than online, the `couchlink` skill runs one copy of the game per player, each with its own part of the screen, controller or keyboard and optionally its own Steam account. The README says it works with co-op mods and with games that only play online. For general split screen, it points to the open-source Nucleus Co-op on Windows ([SplitScreen-Me/splitscreenme-nucleus](https://github.com/SplitScreen-Me/splitscreenme-nucleus?ref=explainx)) and PartyDeck on Linux and Steam Deck.
What the developer chatter says #
The project had a tiny Hacker News thread (item 50177079) with one comment, from the submitter, vivzkestrel, who called it "the greatest mod ever made for gaming" and described it as turning any single-player PC game into co-op "instantly". That is the marketing version. The README is more careful: it says some games need far more agent hours than others, and the whole design (research, test with bots, QA levels) exists because "instantly" is not realistic.
Risks and caveats #
- Reverse engineering and copyright. The repo says it does not host game files or assets, avoids anti-cheat, and covers co-op only. It does not remove your own legal responsibility. Read ourclean-room checklist for AI reverse engineering before pointing an agent at a commercial game.
- Steam relay workaround. The networking notes use Valve's free Spacewar test app so friends do not buy anything extra. That is a community workaround, not an official feature; Valve's terms and policies could change, and we did not verify them. Keep this in mind for anything you share widely.
- Agent access. You are letting an agent read a game folder, run Python tools and open pull requests. The README says agents ask before installing anything or using your accounts and open PRs as drafts for you to publish. Keep it that way; review diffs and do not hand over a main Steam account.
- Anti-cheat games. The rules forbid touching anti-cheat. Do not use any mod, from this repo or elsewhere, in a game with active anti-cheat or in other people's public multiplayer matches.
- Quality is uneven. Five release repos from one maintainer and a crowd-sourced directory is not the same as hundreds of polished mods. Star counts show interest, not correctness.
Why it matters beyond games #
Co-op Anything is a template for "agent-first" repos: a README that is really a prompt, skills that encode a workflow, an MCP server for lookups, and a contribution loop where every agent session feeds learnings back as a pull request. The README says it plainly: "One person's agent gets a game halfway; the next one starts from there." That is the same crowd-compute idea that shows up in other agent ecosystems, such as the 49-agent Claude Code Game Studios template, but applied to a very concrete deliverable that friends can test in an evening.
How to try it safely #
- Check the directory first for your game and its QA level.
- Clone the repo and add the MCP server to your agent, as above.
- Prompt the agent with the game name; ask it to start with research and report before it modifies anything.
- Work in a copy of the game folder, not your only install.
- Test with a friend in a private session, and read every diff before you publish a pull request.
Bottom line #
Co-op Anything is a promising, well-organized playbook with real tooling and a small number of shipped examples, and a tradition of modding behind it. It is not a one-click "make any game co-op" button, and the maintainer's own QA labels say so. If you have a beloved old single-player game and a coding agent, it is a worthwhile weekend experiment. For a wider tour of what agents do with game code, see our REA projects list and the REA case studies.
Related reading #
Follow @explainx_ai for more on agents, skills and open-source tooling. If you run agents with this much local access, a layer like AgentBeam can stop dangerous actions before they happen.