What is human-in-the-loop for AI agents?

Human-in-the-loop for AI agents means pausing the agent at a risky step and letting a real person approve, deny, or answer before it continues. Pushary is the managed layer that does it, built so your agent can ask its own end-users. One tool call sends the question to a person's phone, blocks until they decide, fails closed if nobody answers in your window, and writes the outcome to a durable decisions ledger.

How can my agent ask my own users instead of just me?

You reference each user by your own externalId. Call enroll(externalId) to connect their phone with a one-tap link, then call decisions.ask with the same externalId. The approval lands with that specific person, so a product with thousands of users can ask each of them about their own decision.

What happens if nobody answers?

The decision fails closed. If nobody answers in the window you set, it comes back denied, and you choose what to do. Because the wait is backed by a durable ledger, a restart does not lose the request.

Do my end-users need an account or app?

No. One tap connects their phone. Delivery uses native app push when they have it, PWA push notifications otherwise, and Slack if you route it there. They never see a Pushary login or bill.

Which frameworks does Pushary support?

The Vercel AI SDK and Eve have published adapters (@pushary/ai-sdk and @pushary/eve). LangGraph, LangChain, CrewAI, Mastra, the OpenAI Agents SDK, Hermes, and OpenClaw use @pushary/server in TypeScript or pip install pushary in Python. There is also a hosted MCP server for the Claude Agent SDK and any MCP client.

How much does it cost?

The Partner plan is $99 a month for companies embedding human-in-the-loop for their end-users. There is a $9.99 a month Agent plan for a solo developer who wants their own agent to notify or ask them. Card details are collected at checkout.

Human-in-the-loop for AI agents

Human-in-the-loop for your AI agent's users.

When your agent needs a person to approve or answer, Pushary sends the question to their phone and waits for the decision. Your users connect with one tap. No account, no app, no key. One tool call sends the question, blocks until they decide, fails closed if nobody answers, and writes the outcome to a durable ledger.

Read the two-call quickstart

npm i @pushary/servertwo calls

import { createPusharyServer } from "@pushary/server"

const px = createPusharyServer({ apiKey: process.env.PUSHARY_API_KEY! })

// Connect an end-user's phone once. Keyless, one tap, no app to install.
await px.enroll(user.id)

// Ask a real human. Blocks until they decide. Fail-closed on timeout.
const decision = await px.decisions.ask({
  externalId: user.id,
  type: "confirm",
  question: "Approve a $480 refund to customer #4471?",
})

if (decision.approved) await issueRefund()

Every framework

Add it to the agent you already built

The same two calls, in the idiom of your framework. Published adapters where they help, and the server or Python SDK everywhere else.

Vercel AI SDK

Ask a human from your Vercel AI SDK agent, answered on a phone.

npm i @pushary/ai-sdk

LangGraph

Ask a human from your LangGraph agent, answered on a phone.

pip install pushary

CrewAI

Ask a human from your CrewAI agent, answered on a phone.

pip install pushary

OpenAI Agents SDK

Ask a human from your OpenAI Agents SDK agent, answered on a phone.

pip install pushary

Eve

Ask a human from your Eve agent, answered on a phone.

npm i @pushary/eve

Mastra

Ask a human from your Mastra agent, answered on a phone.

npm i @pushary/server

LangChain

Ask a human from your LangChain agent, answered on a phone.

pip install pushary

MCP

Ask a human from your MCP agent, answered on a phone.

No install. One MCP server entry.

Claude Agent SDK

Ask a human from your Claude Agent SDK agent, answered on a phone.

No install. One mcpServers entry.

Hermes

Ask a human from your Hermes agent, answered on a phone.

pip install pushary

OpenClaw

Ask a human from your OpenClaw agent, answered on a phone.

npm i @pushary/server

On a framework not listed here? The same two calls work anywhere @pushary/server or pip install pushary runs.

How it works

From ask to resume

Connect the phone once, ask a human where a wrong autonomous action has a cost, and act on the decision. The waiting, the delivery, and the record are handled for you.

01
Connect the phone that approves
Call enroll(externalId) with your own id for the user. They tap a one-tap link once to connect their phone. No account, no app, no key on their side.
02
Install the SDK
Use @pushary/server in TypeScript or pip install pushary in Python. On the Vercel AI SDK and Eve, use the @pushary/ai-sdk and @pushary/eve adapters.
03
Ask a human where it matters
Call decisions.ask inside the tool or step that guards the risky action. It sends the question to the person's phone and blocks until they answer.
04
Act on the fail-closed decision
If nobody answers in your window, the decision comes back denied, so a timeout never turns into an unapproved action. Every answer is written to a durable ledger.

A decision is a yes or no confirm, a multiple-choice select, or a free-text input. Every one is stored with its outcome and kept in the audit log.

FAQ

Questions, answered

Give your users the final say, on their phone.

Connect a user with one tap, ask a human with one call, and act on a fail-closed decision. The durable store and the audit log come with it.

Partner plan, $99 a month. Your users never see a Pushary login or bill.