Vercel AI SDK

Gate every Vercel AI SDK tool call through LatchGate’s enforcement pipeline.

Prerequisites

• LatchGate running (latchgate up or production deployment)

Install

npm install @latchgate/ai-sdk latchgate ai

Quick start

import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
import { latchgateToolset } from "@latchgate/ai-sdk";

const { tools, close } = await latchgateToolset({
  gateUrl: "http://localhost:3000",
});

try {
  const { text } = await generateText({
    model: openai("gpt-4o"),
    tools,
    maxSteps: 5,
    prompt: "Fetch https://httpbin.org/get and summarize the response",
  });

  console.log(text);
} finally {
  await close();
}

How it works

latchgateToolset() discovers all registered LatchGate actions via the REST API and returns a { tools, actionIds, close } object. tools is a record of AI SDK tool objects with JSON Schema definitions from each action’s manifest (via the AI SDK’s jsonSchema() helper). When the LLM calls a tool, the wrapper routes it through LatchGateClient.execute() — the full enforcement pipeline runs transparently.

Call close() when the toolset is no longer needed to release the underlying transport. If you provided a pre-configured client via options, close() is a no-op — the client lifecycle remains yours.

Filtering actions

const { tools, close } = await latchgateToolset({
  gateUrl: "http://localhost:3000",
  include: new Set(["http_fetch", "github_read"]),
  exclude: new Set(["send_message"]),
});

UDS transport (production)

Create a LatchGateClient with UDS transport and pass it to the toolset:

import { LatchGateClient } from "latchgate";
import { latchgateToolset } from "@latchgate/ai-sdk";

const client = new LatchGateClient({
  socket: "/run/latchgate/gate.sock",
  agentId: "my-ai-sdk-agent",
});

const { tools, close } = await latchgateToolset({ client });

Alternatively, set the LATCHGATE_URL environment variable and omit gateUrl:

// With LATCHGATE_URL=http://localhost:3000 set:
const { tools, close } = await latchgateToolset();

Audit callback

Use onAudit for programmatic access to receipt metadata after each successful execution:

const { tools, close } = await latchgateToolset({
  gateUrl: "http://localhost:3000",
  onAudit: (record) => {
    // record.actionId, record.receiptId, record.traceId, record.verification
    auditLog.store(record);
  },
});

Discovery timeout

Control the HTTP timeout for the discovery phase (default 15000ms):

const { tools, close } = await latchgateToolset({
  gateUrl: "http://localhost:3000",
  timeout: 10000,
});

More details

Full API reference: latchgate-integrations/ai-sdk/README.md.