AI SDK

Give a Vercel AI SDK agent the looot tools with @looot/ai-sdk. Install, token, a generateText example and the five tools.

@looot/ai-sdk gives an agent built with the Vercel AI SDK five looot tools. With one agent token and one prepaid balance, the agent can search 4,200+ data operations from 300+ providers, such as work emails, phone numbers, company and people search, Google results, web pages and LinkedIn profiles. It reads the price before it runs and pays per call. A failed call costs nothing.

Install

npm install @looot/ai-sdk ai zod

The package needs Node 22.18 or newer. Its peer dependencies are ai 6 or 7 and zod 3.25.76+ or 4.1.8+.

Token

Create an agent token on looot.ai under Settings, Agent tokens. Tick catalog.read, runs.read, runs.execute and usage.read. Export it as LOOOT_TOKEN:

export LOOOT_TOKEN="your-agent-token"

The tools read LOOOT_TOKEN on each call. To pass the token in code instead, use loootTools({ apiKey }). For more on tokens and scopes, see Headless and CI.

Top up the balance before the first paid run. See Money.

Example

import { generateText, isStepCount } from "ai";
import { loootTools } from "@looot/ai-sdk";

const { text } = await generateText({
  model: "openai/gpt-5-mini",
  tools: loootTools(),
  stopWhen: isStepCount(8),
  prompt: "Find the work email of Jane Doe at example.com and tell me what it cost.",
});

console.log(text);

The model string goes through the Vercel AI Gateway. Any AI SDK model provider works the same way. On ai 6, the stop condition is stepCountIs(8).

A typical run takes four tool calls. The model calls searchCatalog with the task in plain words, inspectOperation on the endpoint it picks to read the inputs and the price, runOperation with those inputs, and getRun if the result was not ready within the wait. streamText takes the same tools.

Tools

Tool	What it does	Cost	API call
`searchCatalog`	Ranked endpoints for a task in plain words, with each endpoint’s ID and price.	Free	`GET /v1/catalog/search`
`inspectOperation`	The input schema, output schema and price of one endpoint.	Free	`GET /v1/operations/{endpointId}`
`runOperation`	Runs one endpoint and pays from the balance. Waits up to `waitSeconds` (default 20, max 60), then returns a run ID.	Paid	`POST /v1/runs`
`getRun`	The status, result and cost of one run.	Free	`GET /v1/runs/{runId}`
`getBalance`	The available and reserved balance, and the minimum top-up.	Free	`GET /v1/balance`

loootTools() returns all five. Each one is also exported on its own, for example tools: { searchCatalog: searchCatalog(), runOperation: runOperation() }.

Options

Every tool function and loootTools() take the same options.

Option	Default	Meaning
`apiKey`	`process.env.LOOOT_TOKEN`	The agent token.
`baseUrl`	`https://api.looot.ai`	The looot API address.
`fetch`	`globalThis.fetch`	A custom fetch, for tests or a proxy.
`timeoutMs`	`75000`	Timeout per request, long enough for a 60 second wait.
`requireApproval`	`false`	`runOperation` and `loootTools()` only. When `true`, the AI SDK asks for approval before each paid run.

Results and errors

Read status in the runOperation result. A resolved call is not a success: a bad input comes back with status: "failed" and is not charged. queued or running means the model should call getRun with the run ID.

runOperation sends an idempotency key on every run. The model can pass its own idempotencyKey; the same key with the same input never pays twice. See Idempotency.

When the API answers with a non-2xx status, the tool throws a LoootApiError with status, code and body. A missing token throws missing_token before any request.