Actions & tools

An action is a single operation on a provider — github__list-repos, slack__send-message. The catalog describes every action with a name, a description, and a JSON-Schema for its input, so it can be handed straight to an LLM as a tool.

Naming: provider__action

Tool names join the provider and the action with a double underscore (__):

github__list-repos
slack__send-message

This is the same convention the MCP server and the SDK tool adapters use. (In policies and API-key grants, actions are written with a slash — github/list-repos — to support patterns like github/* or */send*.)

Browse the catalog

GET /api/v1/catalog                  # paginated list of providers (filter by category, tier)
GET /api/v1/catalog/{provider}       # full detail incl. every action's input schema
GET /api/v1/actions/{provider}/{action_name}   # one action's schema + metadata

A catalog action looks like:

{
  "name": "list-repos",
  "description": "List user repositories",
  "http_method": "GET",
  "input_schema": {
    "type": "object",
    "properties": { "per_page": { "type": "integer", "default": 30 } }
  }
}

Risk classification

Every action is classified read, write, or destructive (from its HTTP method and name). This drives default-deny on writes and lets you write policies that target, say, only destructive actions. See Policies.

Calling actions

• Directly: execute / execute-async.
• As LLM tools: agent frameworks.
• Over MCP.