The runtime, documented.

WorldMind exposes three endpoints, one SDK per supported engine, and a mod bridge for modifiable games. Your engine speaks state to the runtime. The runtime speaks structured actions back. Nothing else.

Five steps, end to end. The curl examples assume you have an API key and a project. SDK snippets follow the same sequence.

1. 01
   Create a project

   Projects scope your agents, capabilities, and traces. Create one per game or per environment.

   curl -X POST https://api.worldmind.dev/v1/projects \
     -H "Authorization: Bearer $WM_KEY" \
     -d '{ "name": "ironwatch-mmo" }'

2. 02
   Declare capabilities

   List the actions the engine allows. The validator rejects anything not declared here.

   curl -X POST https://api.worldmind.dev/v1/capabilities \
     -H "Authorization: Bearer $WM_KEY" \
     -d '{ "action": "speak", "args_schema": { "text": "string", "tone": "string" } }'

3. 03
   Register an agent

   Personas, memory profile, capability allow-list, doctrine.

   curl -X POST https://api.worldmind.dev/v1/agents \
     -H "Authorization: Bearer $WM_KEY" \
     -d '{
       "agent_id": "blacksmith_01",
       "archetype": "quest_giver",
       "persona": "gruff but warm village blacksmith",
       "capabilities": ["speak", "give_quest", "complete_quest"],
       "memory_profile": "npc_social_basic"
     }'

4. 04
   Push game events

   Observations flow into the agent's context. Send events as they happen.

   curl -X POST https://api.worldmind.dev/v1/events \
     -H "Authorization: Bearer $WM_KEY" \
     -d '{
       "agent_id": "blacksmith_01",
       "event_type": "player_nearby",
       "payload": { "player_id": "adam", "quest_state": "ready_to_turn_in", "distance": 3.4 }
     }'

5. 05
   Request a decision

   Ask for a structured action. The response is schema-locked and always carries a fallback.

   curl -X POST https://api.worldmind.dev/v1/decisions \
     -H "Authorization: Bearer $WM_KEY" \
     -d '{
       "agent_id": "blacksmith_01",
       "mode": "social",
       "constraints": { "max_latency_ms": 1500, "max_tokens": 1200 }
     }'

An agent is a runtime-controlled entity with cognition. Agents own persona, doctrine, permissions, memory, and an update cadence. A single agent can represent a quest giver, a patrol unit, a colonist, or a director.

{
  "agent_id": "blacksmith_01",
  "archetype": "quest_giver",
  "persona": "gruff but warm village blacksmith",
  "capabilities": ["speak", "give_quest", "complete_quest"],
  "memory_profile": "npc_social_basic",
  "doctrine": ["lore.ironwatch", "role.merchant"],
  "update_schedule": "on_event",
  "fallback_policy": "canned_line"
}

A capability is an action family your engine exposes. Actions are declared with an argument schema. The validator rejects any action the agent is not permitted to emit.

[
  { "action": "speak",             "args": { "text": "string", "tone": "string" } },
  { "action": "move_to",           "args": { "x": "f32", "y": "f32", "z": "f32" } },
  { "action": "cast_skill",        "args": { "skill_id": "string", "target_id": "string" } },
  { "action": "request_backup",    "args": { "location": "string" } },
  { "action": "set_speed",         "args": { "value": "f32" } }
]

Observations are the normalized inputs your game sends into the runtime. They can include entity state, combat events, chat lines, UI snapshots, minimap crops, and screenshots.

{
  "agent_id": "blacksmith_01",
  "event_type": "player_nearby",
  "timestamp": 1712594433.21,
  "payload": {
    "player_id": "adam",
    "quest_state": "ready_to_turn_in",
    "distance": 3.4,
    "reputation": 42,
    "faction_matches": ["ironwatch"]
  }
}

An intent is the high-level decision. A plan is an ordered series of action proposals with rationale and a fallback. The engine accepts, partially accepts, rejects, or requests a repair.

{
  "intent": "greet_and_prompt_turnin",
  "confidence": 0.91,
  "goal": "move_player_toward_quest_completion",
  "next_actions": [
    {
      "action": "speak",
      "args": { "text": "Hey Adam — did you manage to get those iron bars?" },
      "why": "player is ready_to_turn_in; warm opener fits persona"
    }
  ],
  "memory_writes": [],
  "fallback_if_unavailable": "canned_greeting"
}

Memory is structured, not a chat log. Five layers (episodic, semantic, relational, procedural, scratchpad) with salience, decay, and scope. See the memory section on the home page for the layer breakdown.

The primary endpoint. Given an agent and a mode, the runtime builds context, retrieves memory, plans, validates, and returns a structured decision.

{
  "agent_id": "patrol_cruiser_17",
  "mode": "tactical",
  "observations": [
    { "event_type": "pursuit_ongoing", "target": "suspect_vehicle_9" }
  ],
  "constraints": {
    "max_latency_ms": 1500,
    "max_tokens": 1800,
    "allowed_models": ["mid", "strong"]
  }
}

{
  "intent": "intercept",
  "confidence": 0.82,
  "goal": "stop_suspect_without_collision",
  "next_actions": [
    { "action": "navigate_to_road_segment", "args": { "segment_id": "A12" } },
    { "action": "request_backup",           "args": { "location": "B7" } }
  ],
  "memory_writes": [
    { "type": "episodic", "text": "third escape attempt this week" }
  ],
  "fallback_if_unavailable": "continue_safe_pursuit",
  "trace_id": "trc_01HWM3ZNK..."
}

First-party SDKs are available for C# (Unity), C++ (Unreal), TypeScript (hosted), and Python (tooling). All speak the same wire format and share a common schema for capabilities.

For games without engine access, a local bridge process runs alongside the game. Official bridges ship for BeamNG, RimWorld, and Skyrim. The protocol is documented so additional bridges are straightforward.