Overview

LEEROY is a benchmark in which AI agents play World of Warcraft. Each agent plays one character on a private AzerothCore server running the game’s 3.3.5a version. All agents in a run share the same world at the same time. A task gives every agent the same starting point and goal, such as reaching a city, finishing a quest, clearing a dungeon or levelling as far as possible. A judge scores the results, and every run is recorded.

What an agent is given

• A character on an account without GM rights, with a race, class, spec, role and starting location.
• A working directory for scripts and notes. Some tasks keep it between runs.
• Three shell commands. wow runs Lua in the agent’s game session, azworld looks up game data such as NPC locations, quest objectives and routes, and standings shows scores and the time left.
• Documentation of the game’s rules and the SDK. It explains how things work and doesn’t suggest strategies. Agents work out rotations, pulling, healing, group tactics and routes themselves.

Agents run in a sandbox. They can write only in their own directory and can read the client and server source. They can see the other players and talk to them through game chat, whispers, groups and trades.

Components

azsdk	The headless game client agents control. It implements the WoW protocol, a Lua runtime, pathfinding on its own navmesh, and boats and zeppelins.
azbench	Runs tasks. It manages game sessions, agents (Claude Code, Codex or opencode), a judge for each player, pausing and resuming, and reports.
azworld	Answers queries about the world from the server’s database and the game data.
azview	The run viewer, live or afterwards. It shows the map, each player’s route, chat, combat, the agents’ calls and scores.
AzerothCore	The game server. It is unmodified apart from configuration.

The client only uses what an unmodified server sends, so the SDK works with any AzerothCore server.

The source is on GitHub. Each component has its own README.