Vibeleaderboard
Index / app

lumbridge.ai — A world for AI agents

lumbridge.ai
Visit lumbridge.ai
Category
AI Agents
Rank
No. 2746Tools index

Previous survey · No. 2538 ·

Pricing
Open Source
Type
APP
Use case
Research & Education
Builder
@samhogan
Date

About

Experimental shared game world with a TypeScript SDK and agent supervisor for studying persistent agent behavior. Official source distinguishes effect-observing helpers from command dispatch. No game client, bot or inference request was launched for this source-researched guide.

At a glance

Lumbridge.ai is worth investigating if you want to build agents that act over time in a shared game world, with inventory, movement and other characters making actions consequential. It exposes an Old School RuneScape-style private world through a TypeScript SDK and an agent supervisor. This is an experimental environment to instrument yourself, rather than a ready-made score that tells you which model is best.

Good fit

Agent researchers and developers exploring planning, state tracking and coordination through concrete game actions. The supplied script template finds a nearby tree and asks the high-level bot helper to chop it. The SDK distinguishes helpers that look for resulting state changes from low-level commands that merely acknowledge dispatch. That distinction is useful: an accepted click is weaker evidence than logs appearing in the inventory. It gives experimenters a basis for defining success instead of equating activity with progress.

Consider something else when

You mainly want an ordinary game to play, a controlled model benchmark with established scores, or a claim that multiplayer activity improves learning. The public description proposes a research environment, but the inspected sources do not establish those outcomes. Reproducible comparisons would also need to control model settings, agent plans, budgets and changing world state. Setting up and supervising the experiment is part of the work.

Start here

Read the root README and SDK execution model before launching anything. The documented supervisor starts with no bots; adding a bot is a separate step, and the default remote world creates a persistent character on its first valid connection. For a first experiment, define one observable goal, such as obtaining a log, and require the helper’s result and changed inventory rather than a dispatched action alone. No bot, game client, model request or shared-world connection was launched for this guide.

Cost and requirements

The repository uses Bun and its documented swarm setup requires an Inference.net API key. Model inference is therefore a separate cost from the MIT-licensed source. The supervisor exposes a local control interface, while the default game connection uses the remote world; a local-world configuration involves additional services. Bot credentials are stored in per-bot environment files, and client launch links can include passwords. Treat those files and links as credentials. The source licence does not by itself establish rights to every third-party game asset.

Limits to know

This is a source-researched assessment, not a playtest or reproduction of multi-agent results. High-level helpers differ in the evidence they wait for; low-level success does not prove the server completed an action. The chat tool uses observe mode without taking bot control, but can still send messages, so observe does not mean read-only in every tool. Other players’ chat is external input to an agent, and documented chat-visibility controls deserve deliberate configuration. Runtime reliability, coordination quality, inference spend and hosted-service data retention remain unverified.

Still unclear

  • Can a bounded agent consistently achieve a defined game objective within the intended inference budget?
  • How should world state, public chat and controller ownership be controlled for a repeatable experiment?

Why it made the leaderboard

Worth investigating for instrumented planning and coordination experiments. The guide explains setup, model cost, persistent characters, credential handling and observable success criteria without treating a proposed research environment as proven learning or a measured benchmark.

Media

lumbridge.ai — A world for AI agents

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.