Execute one intent

Public documentation preview — incomplete. The SDK and optional AI SDK peers remain private and unpublished. This illustrative flow is not a tested live-provider or production authorization recipe.

After configuring models, servers and policy, router.execute accepts a prompt, trusted context and selected sources. It performs at most one governed tool call. It is not an autonomous planner, transaction or general multi-tool loop.

const context = await getAuthenticatedContext();
if (context.role !== 'reader' ||
    typeof context.customerId !== 'string' || !context.customerId) {
  throw new Error('Authenticated reader and customer ID required');
}

const router = createAppRouter();
try {
  const result = await router.execute({
    prompt: `Get the balance for customer ${context.customerId}.`,
    tools: mcpServers.filter(server => server.serverId === 'billing-server'),
    context,
    signal: AbortSignal.timeout(30_000),
  });
  console.log({ status: result.status, sideEffectState: result.sideEffectState });
} finally {
  await router.close();
}

This integration fragment refers to your host-supplied getAuthenticatedContext, and to createAppRouter/mcpServers from the configuration guide. The source-controlled examples/code.ts module is a structural reference for repository maintainers, not proof of a live provider result. Within this scope the policy makes only get_customer_balance eligible and its exact binding selects billing.routing. The model may propose tool arguments; the router validates them and checks the authenticated customer ID again before dispatch. Host policy must enforce tenant/resource restrictions for your application. Passing the mixed-model server array without narrowing can yield needs_input before ranking or model inference; limit: 1 cannot choose a specialist. Model-access restrictions may instead produce model_access_denied first. An explicit allowed model override is possible but must pass access checks. Inspect the returned status and sideEffectState: denial, approval, missing input, stale binding, timeout, and unknown post-dispatch effects are not success. Do not automatically retry unknown effects or bypass approval.

Get started

Integrate

Guides

Reference

Execute one intent

Execute one intent

​Execute one intent

Execute one intent