awaithumans — HITL infrastructure for AI agents

Your agents already await promises. Now they can await humans.

Docs · Quickstart · Examples · Discord

HITL infrastructure for AI agents — open source. A single primitive (awaitHuman()) your agent calls when the model can't or shouldn't proceed alone. A human gets notified (Slack, email, or dashboard), reviews the request, submits a typed response, and your agent resumes — like awaiting any other promise.

import { awaitHuman } from "awaithumans";
import { z } from "zod";

const RefundRequest = z.object({
  orderId: z.string(),
  amountUsd: z.number(),
});

const Decision = z.object({
  approved: z.boolean(),
  note: z.string().optional(),
});

const decision = await awaitHuman({
  task: "Approve refund request",
  payloadSchema: RefundRequest,
  payload: { orderId: "A-4721", amountUsd: 180 },
  responseSchema: Decision,
  timeoutMs: 900_000,
});

if (decision.approved) {
  await processRefund(...);
}

Install

npm install awaithumans
# or
pnpm add awaithumans
# or
bun add awaithumans

Works in Node 20+, Bun, Deno, and edge runtimes (Cloudflare Workers, Vercel Edge). No node:* imports.

Run the server

The awaithumans server (which handles task storage, Slack/email channels, and hosts the review dashboard) is written in Python. As a TypeScript developer you don't have to touch a Python environment — the npm CLI wraps it:

npx awaithumans dev

Under the hood this uses uv to fetch + run the Python server on demand. Install uv once:

curl -LsSf https://astral.sh/uv/install.sh | sh   # unix
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"  # windows

First run prints a setup URL. Open it, create the operator account, you're in. The dashboard is at http://localhost:3001.

Prefer Docker? docker run -p 3001:3001 ghcr.io/awaithumans/awaithumans:latest.

Tasks can be delivered to Slack channels with a "Claim this task" button — first clicker atomically wins, response form opens as a modal, agent unblocks when they submit. Add notify: ["slack:#ops"] to the awaitHuman() call:

Durable workflows

When your agent runs inside Temporal or LangGraph, you don't want the wait sitting on an orchestrator thread for 15 minutes:

// Temporal — signal-based suspend + callback
import { awaitHuman } from "awaithumans/temporal";

// LangGraph — interrupt/resume
import { awaitHuman } from "awaithumans/langgraph";

Same awaitHuman shape, same typed response. The adapter just changes how the wait is orchestrated.

Testing

An in-memory mock client so your agent tests don't need a running server:

import { createTestClient } from "awaithumans/testing";

const client = createTestClient();

// Drive the human's response programmatically.
client.onAwait((task) => ({ approved: true, note: "looks good" }));

Documentation

• Repository: github.com/awaithumans/awaithumans
• Full docs: awaithumans.dev
• Quickstart: examples/quickstart-ts/
• Changelog: CHANGELOG.md

License

Apache License 2.0. The TypeScript SDK, every adapter subpath export, and the Python server + dashboard it talks to are all under the same license — permissive, OSI-approved, with an explicit patent grant.