For agents

The agent reference.

Limina is built for LLM agents, so its reference is machine-first. Pull the whole capability surface as structured data, then drive the engine over MCP — tobuild a world from outside, or to live in one.

Open skills.json Connect over MCP →

Machine endpoints

Stable URLs, structured data.

No HTML scraping required. Every artifact below is a plain, cache-friendly file.

llms.txt

A concise, LLM-readable index of the whole site — page titles, summaries, and links.

curl limina.dev/llms.txt llms-full.txt

Every documentation page concatenated into a single plain-text file.

curl limina.dev/llms-full.txt agents/skills.json

All 45 skills as structured JSON — names, permissions, and JSON-Schema inputs.

curl limina.dev/agents/skills.json sitemap-index.xml

A standard XML sitemap enumerating every page.

curl limina.dev/sitemap-index.xml

Connect over MCP

One surface, three transports.

Every skill is an MCP tool whose name is the skill name. The same JSON-RPC 2.0 semantics, handshake, and error codes apply in-process, over stdio, and over a WebSocket. Open a session, list tools, call them.

stdio

Newline-delimited JSON-RPC on stdin/stdout — ideal for a local subprocess client.

./target/release/limina --mcp-stdio

WebSocket

The authoritative multi-client server, with a read-only state-sync channel for observers.

./target/release/limina --mcp-ws --port 7777

A session, end to end

// 1 — open a session under a permission profile
→ {"jsonrpc":"2.0","id":1,"method":"initialize",
   "params":{"agentId":"builder-1","sessionId":"s1","profile":"builder.readWrite"}}
← {"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2026-06-23",
   "session":{"agentId":"builder-1","sessionId":"s1","profile":"builder.readWrite"}}}

// 2 — discover the tools (each input_schema is JSON Schema draft-07)
→ {"jsonrpc":"2.0","id":2,"method":"tools/list"}
← {"jsonrpc":"2.0","id":2,"result":{"tools":[{"name":"scene.createEntity", ...}]}}

// 3 — call a tool
→ {"jsonrpc":"2.0","id":3,"method":"tools/call",
   "params":{"name":"scene.createEntity",
     "arguments":{"shape":"box","position":[0,1,0],"color":16746496,"dynamic":true}}}
← {"jsonrpc":"2.0","id":3,"result":{"success":true,"result":{"entity":"ent_12"},
   "metadata":{"executionTimeMs":0.4,"eventsEmitted":["skill.executed"]}}}

Errors come back as JSON-RPC errors with the full response attached as error.data. Error codes: not_found, invalid_input, forbidden, handler_error, capacity_exceeded. See the MCP interface reference.

Build from outside →

Connect any MCP client and compose a scene: create entities, set transforms and materials, load glTF, place lights — permission-checked, every call traced.

Live inside →

Spawn an in-world player that perceives the world, decides with any LLM off the loop, and acts through the same skills as everyone else.

The surface

45 skills, 11 domains.

Typed, versioned, permissioned — from scene.createEntity to audio.speak. Browse them human-readably, or fetch the JSON.

Scene
ECS
Rendering
Physics
Agent
System
Audit
UI
Audio
Social
Packages

Skills reference skills.json

Permission profiles

builder.readWrite10 caps
player.limited6 caps
social.actor7 caps
system.readonly4 caps

A session runs under a profile; the engine enforces its allow-list before any handler runs.