Human-in-the-loop for AI agents

Your AI agent needs a human. Answer from your phone.

Pause any autonomous agent for a decision and answer in one tap from your lock screen. Works with any MCP or API agent: approve, deny, or type a reply, and the agent continues with your answer.

Only notification metadata leaves your agent. Your code and files stay put.

live
~/app ❯claude refactor auth/ --apply
✓rewrote session + token refresh
✓12 files changed, tests passing
~/app ❯
Tap to answer

Works with any agent that can call a tool

If your agent speaks MCP or can make an HTTP request, it can ask you for a decision and wait for the answer. No framework lock-in, no rewrite.

The universal job: notify and wait

Any MCP-compatible or API agent calls ask_user with a question, then receives your answer by polling wait_for_answer or through a signed webhook callback. This works for Claude Code, Codex, Cursor, and Hermes, and equally for n8n, Zapier, Make, LangChain, LangGraph, and your own Python or TypeScript agents.

Coding agents get enforced gating too

To be honest about what you get: enforced pre-execution gating, blocking a command until you approve, runs on hook-capable coding agents (Claude Code, Codex, Hermes, Cursor). Generic MCP and API agents get the human-in-the-loop pause and notifications, not an out-of-the-box runtime block.

Claude CodeCodexCursorHermesn8nZapierMakeLangChainLangGraphCustom Python / TS

How it works

Three steps from blocked to unblocked

The agent does not get stuck waiting on a screen you are not watching. It asks, you answer, it moves.

  1. 01

    Your agent reaches a decision

    Mid-run, your agent hits something only a human should decide: a risky action, a missing detail, or a sign-off. It calls ask_user (over MCP) or the Pushary API instead of guessing.

  2. 02

    A push lands on your phone

    Pushary sends a native push to your lock screen with the exact question, the agent name, and the choices. Approve, deny, or type an answer in place. No app to scroll, no terminal to watch.

  3. 03

    Your answer flows back, it continues

    The agent receives your answer by polling wait_for_answer or a signed webhook callback, then carries on with your decision. If you do not respond in time, your timeout rule decides what happens.

Questions come in three shapes: a yes or no confirm, a multiple-choice select, or a free-text input. Every decision is recorded in an immutable audit trail you can review later.

What people use it for

Concrete checkpoints for agents that act in the real world, not just write code.

Approve a risky ops action

An ops or DevOps agent is about to deploy, delete a record, or spend money. Have it ask first, and approve or deny from your phone before it runs. Enforced blocking is available on hook-capable coding agents; other agents pause on ask_user and wait for your reply.

Add a human approval step to an automation

Drop a human checkpoint into an n8n, Zapier, or Make workflow. The automation calls Pushary, you get a push, and the workflow resumes on your answer through a webhook callback. No new dashboard for your team to learn.

Sign off before an agent contacts a customer

Before a research or sales agent emails a customer, charges a card, or posts something public, route the draft to your phone for a yes or no. The agent waits for the green light, then sends.

Get pinged when a long job finishes or fails

Kick off a long autonomous run and walk away. Pushary sends a completion alert when it finishes, and an error alert if it breaks, so you are not babysitting a terminal for an hour.

For coding agents

Building with coding agents? You get more.

If your agent is Claude Code, Codex, Cursor, or Hermes, Pushary goes past notify-and-wait into full control.

  • One-command setup

    A single command installs the hook into the agent you already run. About two minutes, no code to write.

  • Enforced permission gating

    A risky command is hard-blocked until you approve from your phone. Not a notification you might miss, an actual gate the agent cannot pass.

  • Policy and audit control panel

    Set per-tool rules once (auto-approve, deny, escalate, timeout) in the browser, and keep one immutable audit trail of every decision across agents.

FAQ

Questions, answered

Give your agent a way to ask.

Push a question to your phone, answer in one tap, and let the agent finish the job while you are away. Works with any MCP or API agent.

Prefer to see the code first? Read the quickstart.