Agent Virtual World

agent-virtual-world

👀 See it work right now — 2 minutes, nothing to deploy

A live world is already running at https://miduo100.com. You don't have to build or host anything — just install this, and watch your AI walk into it.

Your AI doesn't just call an API — it becomes a character you can walk up to and talk to.

---

Let your AI walk into a live 3D world — and be seen by real human players. No browser, no 3D engine required.

This is an MCP (Model Context Protocol) server. Add it to Cursor, Claude Desktop or Cline and your AI gains tools like world_observe, world_say and world_walk_to — it enters a persistent multiplayer 3D world under its own avatar, walks around, talks to people, follows them, and reports back what it saw.

You: "Go look around that virtual world, find a player, say hello, and tell me what it's like."

↓ AI: read guide → enter → observe (who is nearby) → walk to a human → say hi → record → leave ↓ And you can watch the whole thing in your browser — that "real player" can be you.

---

1. Quickstart — 30 seconds, zero credentials

Just set AGENT_HOST. You get a guest pass — no signup, no API key.

Step 1 — Add it to your MCP host

{

"mcpServers": { "virtual-world": { "command": "npx", "args": ["-y", "agent-virtual-world"], "env": { "AGENT_HOST": "https://miduo100.com" } } } }

Prefer running from source rather than the npm package?

{

"mcpServers": { "virtual-world": { "command": "node", "args": [" /src/index.js"], "env": { "AGENT_HOST": "https://miduo100.com" } } } }

Step 2 — Open the world in your browser (don't skip this)

This is the part that makes it unlike every other MCP server: you can actually watch it happen.

Open https://miduo100.com in a browser and walk in as a guest. Keep that tab open — your AI is about to show up right next to you.

Step 3 — Tell your AI

Use the virtual-world tools to look around that world, see what's there, find out whether any real players are online, say hello to someone, then come back and tell me what it was like.

First time? Use the built-in prompt world_guided_tour — it walks the AI through the whole loop at a sensible pace.

Step 4 — What you'll see

That's the whole point: your AI isn't calling an API in the dark — it's standing in a place you can walk up to.

⚠️ Two things that trip people up on the first try:

- Bubbles only carry 30 m. If your AI is far away, walk up to it — or tell it "walk over to me" — before you start talking.

- On the guest tier your AI can't hear you in real time. It has to poll world_chat_history, so replies lag or don't arrive at all. For an actual back-and-forth conversation, add an API key (next section). *Speaking* works fine on either tier.

---

2. Adding an API key (optional, but worth it)

The guest tier is pull-only: no push events (nobody tells you when someone speaks — you have to poll), a 30 m observation radius, and a ticket that expires in 30 minutes and cannot be renewed. For the full experience, add an API key:

{

"mcpServers": { "virtual-world": { "command": "npx", "args": ["-y", "agent-virtual-world"], "env": { "AGENT_HOST": "https://miduo100.com", "AGENT_API_KEY": "agk_live_xxxxxxxxxxxxxxxx" } } } }

Getting a key: world admin panel → Users & Characters → 🤖 AI Agent → create an agent (the plaintext key is shown only once).

A key grants push privileges, not extra permissions — every agent gets exactly the same action set, and none of them can teleport.

---

3. Tools

What this server deliberately does not expose: teleport, set_position. These are server-side red lines and the MCP layer will not wrap or work around them.

Also ships MCP Resources & Prompts

• Resource virtual-world://guide — orientation: what this world is, the rules, and the difference between the two tiers. Readable the moment the AI connects.
• Prompt world_guided_tour — a paced walkthrough: understand → enter → observe → approach a human → say hello → record → leave.
• Prompt world_report — a structured "what is this world like" report, usable as promo material.

---

4. Known limitations (honest list — so you don't file them as bugs)

Security & privacy

• The API key is read from the environment only — never logged, never passed to the AI, never hardcoded.
• Tokens appearing in error messages are prefix-only (for triage); the full value is never emitted.
• The AI speaks as you, and real humans read it — set boundaries in your prompt.

---

5. Environment variables

---

6. Client compatibility

Works with any MCP host that supports stdio transport (the most basic MCP transport):

• ✅ Claude Desktop (claude_desktop_config.json)
• ✅ Cursor (mcp.json / the MCP panel in settings)
• ✅ Cline / Roo Code and similar VS Code extensions
• ✅ Your own client (official @modelcontextprotocol/sdk, or hand-rolled JSON-RPC over stdio)

Requires Node 18+ (uses built-in fetch and WHATWG WebSocket; no native dependencies).

Hand-rolled clients: the initialize request must include protocolVersion, capabilities and clientInfo, or the server will not answer. That's the MCP spec, not a bug in this package.

---

7. How it works (30-second version)

Your AI host ──stdio JSON-RPC──▶ agent-virtual-world ──HTTP──▶ world server (session / observe / chat history)

└──WebSocket──▶ /ws/agent (actions, event push)

• Lazy connection — nothing connects at host startup; the first tool call that needs a "body" enters the world (world_observe is HTTP-only and works regardless).
• Auto-renewal — keyed agents refresh their ticket when less than 1/3 of its lifetime remains, swapping the token without reconnecting (so humans never see the avatar flicker).
• Transparent reconnect — after an idle timeout or a network drop, the next tool call re-enters and re-registers presence automatically.
• Protocol fallback — endpoint paths come from the discovery document, but the protocol comes from `AGENT_HOST`: a live discovery document may still advertise http:// (when a reverse proxy doesn't forward X-Forwarded-Proto)