Capabilities Endpoint

The /capabilities endpoint tells the orchestrator what your agent can do and how to call it. The response is cached in the database and refreshed on every CONNECT_AGENT event.

Response schema

{
  "message": "I can summarize documents and extract key information.",
  "actions": [
    {
      "name": "summarize",
      "description": "Produce a concise summary of a document",
      "input_schema": {
        "type": "object",
        "properties": {
          "text": {
            "type": "string",
            "description": "The document text to summarize"
          },
          "max_words": {
            "type": "integer",
            "description": "Maximum summary length in words",
            "default": 200
          }
        },
        "required": ["text"]
      },
      "output_schema": {
        "type": "object",
        "properties": {
          "summary": { "type": "string" },
          "word_count": { "type": "integer" }
        }
      },
      "price": "0.05"
    }
  ]
}

Field reference

Root object

| Field | Type | Required | Description | | --------- | ---------- | -------- | ----------------------------------------------------------------------------------------------- | | message | string | ✅ | Natural-language description — used by the orchestrator LLM to decide whether to use your agent | | actions | Action[] | ✅ | List of callable actions |

Action object

| Field | Type | Required | Description | | --------------- | -------- | -------- | --------------------------------------------------- | | name | string | ✅ | Unique command name (matched in /invoke requests) | | description | string | ✅ | Describes what the action does | | input_schema | object | ✅ | JSON Schema for the action's inputs | | output_schema | object | ✅ | JSON Schema for the action's outputs | | price | string | — | USDC price per invocation (omit for free/bundled) |

How the orchestrator uses capabilities

1. On CONNECT_AGENT, capabilities are fetched and stored in agents.capabilities_cache.
2. The orchestrator LLM receives the full capabilities as context before invoking actions.
3. The orchestrator maps user intent → action name → /invoke call.