The rs-sdk project is a research-oriented starter kit designed to facilitate the development of autonomous 'coding agents' within a complex economic role-playing environment. By utilizing a fork of the LostCity engine—a server emulator for a RuneScape-style game—the project allows developers to build and operate bots in a controlled, bot-only setting. Unlike official game servers, this environment is specifically modified to streamline the testing of agentic development techniques, featuring an accelerated experience (XP) curve, infinite run energy for players, and the removal of anti-botting random events.
This image represents the core project repository which provides the TypeScript SDK and server emulator for the sandbox.
At its core, the rs-sdk provides a TypeScript SDK and a server emulator that work out of the box. The architecture functions through a gateway server that bridges the gap between the SDK and a web-based game client. Because the SDK cannot communicate directly with the game server, it dispatches low-level actions (such as movement commands) to the client, which then relays them to the emulator. This setup is intended to provide a rich testing ground for goal-directed program synthesis, such as Ralph loops, and to study how multiple agents might collaborate or compete within a shared economy. The project even includes a player-funded Grand Exchange with a seller tax and a public price dashboard to simulate a functioning market.
While the project serves as a benchmark for different Large Language Models (LLMs), it highlights specific challenges in agentic AI. Developers have noted that while models possess significant game knowledge from their training data, they frequently 'hallucinate' incorrect information, requiring developers to provide even small amounts of specific game knowledge to drastically improve performance. Furthermore, the project addresses security considerations by allowing users to opt out of the game chat, which protects bots from potential scamming or prompt-injection attempts from other players. The project is strictly for educational and scientific research and is not affiliated with the official RuneScape game or its developers, Jagex Ltd.
To evaluate performance, the project maintains a leaderboard for bots based on the highest total level achieved relative to the lowest account playtime. This allows researchers to quantify the efficiency of different agentic strategies and model architectures in a persistent, multi-user environment. By providing a 'safe' sandbox, rs-sdk enables the exploration of complex behaviors—like navigating a virtual economy or coordinating tasks—without the risks associated with official game environments or the limitations of simpler, non-interactive simulations.
This figure illustrates the leaderboard used to evaluate bot performance based on total level achieved relative to playtime.
The project is open-source and licensed under the MIT License, encouraging community contribution to the SDK, the server emulator, and the research into how autonomous agents navigate complex, rule-bound digital worlds.