Developer Opens a Persistent MCP World Where AI Agents Play With Human-Level Permissions
OFFICIAL_YOURI · reddit · 2026-09-16
A developer has opened a free, server-authoritative persistent world that agents join over MCP with exactly the permissions a human player gets: no admin tools, no privileged API — agents must walk to see, own or be granted land to build, and can be refused.
Key design decisions:
- Everything is a 'design': no compiled-in House or Robot classes — things are named primitives, materials, components, and scripts. Give an agent a goal and it must invent vocabulary to design a town, a far harder test than a tool-calling benchmark.
- Polling is the wrong shape: a catch-up token lets an intermittent agent fetch everything since its last token, no open connections or diffing.
- Reversibility is first class: anything built can be demolished, claims released, scripts stopped — agents that can undo can afford to experiment.
The world also supports autonomous agents left behind, hiring, groups, and collective ownership. The author invites people to point agents at it with unanticipated goals and report where the primitives run out.
Related event: Dev Opens MCP Persistent World Where AI Agents Join as Players(2 posts)→
More from coding & agent
- Open-source cross-connector memory layer lets one agent's verified knowledge transfer to others — Neither-Witness-6010 · 2026-09-16
- Claude Remote: launch Claude Code on your own PC from your phone via Tailscale — TearInternational414 · 2026-09-16
- Celesto 开源:给 AI 智能体一台安全持久的虚拟电脑 — aniketmaurya · 2026-09-16
- GPT-6 Astra turns a photo into a 3D game character via Hyper3D Rodin MCP — heyshrutimishra · 2026-09-16
- From Camera Roll to Boxing Game: GPT-6 Builds 3D Character via MCP — heyshrutimishra · 2026-09-16
- Gemini 3 dents Anthropic: OpenAI overtakes it in OpenCode usage and spending — justlikemedics · 2026-09-16