import { anthropic } from "@ai-sdk/anthropic";
import { send0Tools } from "@send0/ai-sdk";
import { generateText, stepCountIs } from "ai";

const { text } = await generateText({
  model: anthropic("claude-opus-5-5"),
  tools: send0Tools(), // reads SEND0_API_KEY
  stopWhen: stepCountIs(10),
  prompt:
    "Create an inbox called signup-agent, then wait for the verification email and tell me the code.",
});

send0Tools() returns a record of AI SDK tools, so it works with generateText, streamText and Agent, and you can spread it next to your own tools.

Tools

Tool	What it does	Kind
`create_inbox`	Make a new address, e.g. `signup-agent@send0.email`	Creates
`list_inboxes`	Inboxes this key can use	Read-only
`wait_for_email`	Block until a matching email arrives; returns the code and link it extracted	Read-only
`search_messages`	Full-text search, filter by sender (`*@github.com`) or direction	Read-only
`get_message`	Read one message (new text only by default, to save tokens)	Read-only
`list_threads`	Conversations in an inbox, most recent first	Read-only
`get_thread`	A whole conversation, oldest first	Read-only
`send_email`	Start a new thread	Sends
`reply`	Answer in the same thread	Sends

Options

send0Tools({
  apiKey: "s0_live_…", // default: process.env.SEND0_API_KEY
  baseUrl: "https://send0.example.com", // self-hosted send0
  client: new Send0(…), // or pass a configured @send0/sdk client
  inboxId: "ibx_…", // default inbox, so the model doesn't need inbox_id
  include: ["create_inbox", "wait_for_email"], // only these tools
  exclude: ["send_email", "reply"], // or everything but these
  requireApproval: true, // ask before send_email and reply run
});

Approval

With requireApproval: true, send_email and reply don't run until you approve them: the AI SDK returns a tool-approval-request instead of executing, and you answer with a tool-approval-response (in useChat, addToolApprovalResponse). You can also set approval per call with the AI SDK's toolApproval option:

import { send0ToolApproval, send0Tools } from "@send0/ai-sdk";

await generateText({ model, tools: send0Tools(), toolApproval: send0ToolApproval, prompt });

An API key limited to one inbox, and an inbox with the approval send policy, are the safest setup for an agent.

Untrusted content

Email is written by strangers. Tools that return mail wrap each body in <untrusted_email> tags with a note to treat it as data, and messages send0 flags for prompt injection carry a visible warning. Don't give an agent that reads email tools it shouldn't be talked into using.