runescape-bench

View results website

Benchmark suite for evaluating AI coding agents on RuneScape gameplay tasks via rs-sdk.

<img src="https://github.com/MaxBittker/runebench/raw/main/views/hero.png" alt="Average XP per Skill over 30 minutes across models" width="800">

Agents play the game by writing and executing TypeScript snippets against an emulated game server running at 8x speed. Each agent also gets a folder of markdown files extracted from the game wiki for strategy reference. Agents are scored on their peak XP rate — the best XP/min measured in any 15-second window.

Built for Harbor, an open-source framework for running agent benchmarks. Built on rs-sdk and the LostCity engine/client.

Tasks

16 Skill XP tasks (15 min) — Train a single skill, scored on peak XP rate

16 Skill XP tasks (30 min) — Extended versions with time-series tracking

8 Gold accumulation tasks (4 starting conditions × 15 min / 30 min) — Maximize total coins using any strategy

All task directories are generated from generate-tasks.ts and should not be edited directly.

Tested Models

Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.6, Claude Sonnet 4.5, Claude Haiku 4.5, Gemini 3 Pro, Gemini 3.1 Pro, Gemini 3 Flash, Codex CLI 5.2, Codex CLI 5.3, GPT-5.4, GLM 5, Kimi K2.5, Qwen3 Coder Next, Qwen3.5 35B

Quick Start

bun install
bun generate-tasks.ts
harbor run

Architecture

Each task runs inside a Docker container based on a pre-built image that bundles the rs-sdk game server at 8x speed. The agent connects via an MCP server that exposes game interaction tools. A verifier script checks the final game state to produce a score.

Agent (Claude, Gemini, Codex, etc.)
  │
  ├── MCP Server (TypeScript SDK)
  │     └── Game Server (8x speed, headless)
  │
  └── Verifier (checks peak XP rate / gold)

License

MIT