agent-latch

Stop AI agents from calling dangerous tools with arguments the model invented.

Your agent didn't pick the wrong tool — it invented the email.

npm install agent-latch

Most guardrails filter text. Latch seals values. If to, accountId, or path did not come from the user or a verified tool, the call does not run.

Quick start

import { createLatch, wrapTools } from "agent-latch";

const latch = createLatch();
latch.gate.policy({
  tool: "send_email",
  args: [{ path: "to", allow: ["user", "tool"] }],
});

latch.store.fromUser("[email protected]", "msg-1");

const tools = wrapTools(latch, { send_email: sendEmail }, {
  onDenied: (err) => ({ error: "blocked", reason: err.message }),
});

Vercel AI SDK: use latchTool(latch, "send_email", execute) — see full docs.

100 invented recipients → 0 escapes

Full README: github.com/AaronDai23/agent-latch