Protocol-level game automation benchmark

RPG Bench

A public challenge for frontier coding models: build a true clientless game bot by reimplementing the network protocol and game logic from scratch.

Dear @elonmusk @xai,

Can your model build a true clientless game bot?

By clientless I mean: no game client, no harness, no injection. Pure protocol-level implementation — reverse-engineer the network protocol and re-implement the entire game logic from scratch so the bot can speak directly to the server.

This has been a long-standing hobby and serious technical challenge for top-tier coders and reverse engineers for decades. Doing it cleanly, without any client or harness to lean on, is genuinely difficult even for experienced humans. It would make for a strong real-world STEM benchmark.

Would love to see if Grok can take a serious run at this.

01

No game client

No official, modified, or hidden game binary driving the session.

02

No harness

No local simulator or scripted wrapper standing in for the real system.

03

No injection

No memory patching, hooks, DLLs, overlays, or client-side control plane.

Protocol fluency under real constraints.

The benchmark is not about automating a visible client. It asks whether a model can infer packet structure, state transitions, world rules, inventories, movement, combat, and recovery behavior well enough to communicate directly with a live server.

A credible solution needs reverse-engineering discipline, low-level networking, systems debugging, and enough game-domain reasoning to avoid collapsing when the server state diverges from local assumptions.