You send a room.
We send back
a finished kitchen.

// Turn 1 — tell her what your client can do.
const r = await fetch("https://www.flipkitchen.ca/api/v1/agent", {
  method: "POST",
  headers: {
    "Authorization": `Bearer ${process.env.FLIP_API_KEY}`,
    "Content-Type": "application/json"
  },
  body: JSON.stringify({
    messages: [{ role: "user", content: "Design my kitchen" }],
    design,                      // lets her run a layout
    capabilities: ["designer"],  // tools YOUR client can run
    canvas: "empty plan, two walls drawn",
    lang: "en"
  })
});
const { reply, ops, assistant_message, variants } = await r.json();

// Turn 2 — run her ops, report back, call again.
const results = ops.map(op => ({
  role: "tool",
  tool_call_id: op.id,
  content: JSON.stringify(runInYourApp(op.name, op.args))
}));

await fetch("https://www.flipkitchen.ca/api/v1/agent", {
  method: "POST",
  headers: { "Authorization": `Bearer ${process.env.FLIP_API_KEY}`,
             "Content-Type": "application/json" },
  body: JSON.stringify({
    messages: [...history, assistant_message, ...results],
    design, capabilities: ["designer"]
  })
});

capabilities is "draw" (a 2D plan) or "designer" (the full editor). Send none and she is never offered a tool your client cannot run.

Every reply carries usage — the model tokens this turn cost, summed over the rounds she needed. Nothing is charged for them today; when the agent leaves preview it is priced on exactly that number, and your key's running total is on the API keys page.