Tools for agents from an HTTP API's description: one tool per operation, with the API's own sign-in and keys.

{:ok, analytics} =
  WardrikRest.API.load(
    {WardrikRest.Discovery,
     url: "https://analyticsdata.googleapis.com/$discovery/rest?version=v1beta"},
    name: "ga4",
    base_url: "https://analyticsdata.googleapis.com/",
    auth: %{"oauth2" => {:credential, "google"}},
    only: ["properties_runReport", "properties_getMetadata"],
    read_only: ["properties_runReport"]
  )

WardrikCore.start_agent(
  agent_id: "agt_marketing",
  model: "anthropic:claude-haiku-4-5",
  capabilities: [{WardrikRest, apis: [analytics]}],
  credentials: {MyApp.Credentials, []}
)

• Each tool is named <api>__<operation>, as ga4__properties_runReport, and takes the operation's path, query and body arguments.
• GETs and the operations given as :read_only are read-only; the rest need a person's approval unless the policy allows them.
• Answers are labelled "outside" unless the API is :trusted.
• A tool whose schema is over the API's :max_schema_bytes names its deeper types instead of listing their fields, and <api>__describe gives them.

Descriptions are read by a WardrikRest.Reader. See WardrikRest.Discovery and WardrikRest.API.