Top 10 ways to get alerted when your AI agent needs a human
From MCP-native escalation to PagerDuty, Twilio and Slack webhooks: 10 ways an AI agent can reach a person, ranked by how well they get an answer back.
Sooner or later an agent hits something it should not decide alone. What matters then is whether the right person sees the question in time, and whether their answer gets back to the agent. We ranked ten ways to do it on exactly that: reach (does it get to a person who is away from their desk?), escalation (what if they do not answer?) and round trip (does the answer come back to the agent as data?).
| # | Option | Reach | Escalation | Answer back to agent |
|---|---|---|---|---|
| 1 | redphonr | Slack, SMS, WhatsApp, Telegram, Discord, email, voice | Ladder and on-call | Yes, structured |
| 2 | PagerDuty | App push, SMS, phone, email | Policies and schedules | Acknowledge or resolve only |
| 3 | incident.io On-call | App, SMS, phone | Escalation paths | Acknowledge only |
| 4 | Pushary | Mobile push, Mac app, Slack | Fallback rules | Yes |
| 5 | n8n send and wait | Slack, Teams, email, Telegram, WhatsApp | Build it yourself | Yes, inside n8n |
| 6 | Slack app | Slack | No | Yes, if you build it |
| 7 | Twilio | SMS, voice, WhatsApp | Build it yourself | Yes, if you build it |
| 8 | Telegram bot | Telegram | No | Yes, if you build it |
| 9 | ntfy | Push via app | No | Action buttons, if you build it |
| 10 | Jira Service Management | App, SMS, phone, email | Policies and schedules | Acknowledge only |
1. redphonr
redphonr is built for the full round trip. The agent calls an MCP tool such as request_approval or alert_human; redphonr works down your ladder (Slack, then SMS, then a call) and on to the on-call backup; the first answer is returned to the agent as JSON and every other pending notification is cancelled. On a call you press 1 or 2. Deny on timeout is the default.
Pick it when an unattended agent needs a decision, not just an acknowledgement.
2. PagerDuty
The standard for paging engineers about incidents, with mature schedules, escalation policies, and a mobile app. An agent can open an incident through the Events API. What comes back is an incident state (triggered, acknowledged, resolved), not an approve or deny decision. Pricing is per user. See redphonr vs PagerDuty for when each fits.
Pick it when an agent’s alert is really an incident for an engineering team that already lives in PagerDuty.
3. incident.io On-call
A modern alternative to PagerDuty with on-call schedules, escalation paths and paging by app, SMS and phone, tightly integrated with Slack-based incident response. Like PagerDuty, it is designed around incidents rather than agent decisions.
Pick it when your team runs incidents in incident.io and the agent’s alert should open one.
4. Pushary
Sends approval requests to a phone app, a Mac menu-bar app and Slack, with auto-approval rules and fallbacks. Strong for coding agents like Claude Code and Cursor.
Pick it when a single developer wants to approve their coding agent from their phone.
5. n8n send and wait
If your agent is an n8n workflow, the “send and wait for response” operation pauses the flow until someone answers in Slack, Teams, email, Telegram, WhatsApp and other apps.
Pick it when you already build in n8n. Escalation and on-call are extra nodes you maintain.
6. A Slack app with buttons
A Slack app can post a message with approve and deny buttons and receive the click on your server. Note that plain incoming webhooks can post but cannot receive button clicks; you need a Slack app with interactivity.
Pick it when every approver lives in Slack during working hours, and you are happy to host the callback.
7. Twilio SMS and voice
Twilio gives you the raw pieces: send an SMS, receive the reply on a webhook, place a call and gather a keypress. Everything else (state, retries, escalation, local numbers, audit) is yours.
Pick it when you want full control and have time to build and run it.
8. A Telegram bot
Telegram bots are free and support inline keyboard buttons. Popular with solo developers and small teams.
Pick it when your approvers use Telegram and the stakes are low.
9. ntfy
ntfy is an open-source, self-hostable push notification service. Any script can curl a message to a topic, and notifications can carry action buttons that call a URL.
Pick it when you want simple, self-hosted push notifications for yourself.
10. Jira Service Management
Atlassian’s on-call and alerting now lives in Jira Service Management and Compass, which replace Opsgenie (support ends 5 April 2027). Similar to PagerDuty in shape.
Pick it when you are an Atlassian shop. If you are leaving Opsgenie, read our Opsgenie alternative guide.
The short version
If the agent only needs to tell someone, any of these work. If it needs to ask someone and wait for the answer, you need the round trip: redphonr, Pushary or n8n out of the box, or Slack, Twilio and Telegram if you build it yourself.
Frequently asked questions
How can an AI agent notify a human?
The agent calls a tool or API that sends a message: an MCP server like redphonr, an incident tool like PagerDuty, a Slack or Telegram bot, an SMS through Twilio, or a push notification through ntfy or Pushary. If the agent needs an answer, choose an option that returns the response to the agent, not just a one-way alert.
What is the difference between alerting and human-in-the-loop?
Alerting tells a person something happened. Human-in-the-loop asks a person a question and waits for the answer before continuing. Paging tools are built for the first; approval tools like redphonr handle both.
Can an AI agent call me on the phone?
Yes. redphonr places voice calls where you press 1 to approve and 2 to deny. You can also build calls yourself with Twilio's voice API, or use PagerDuty or incident.io, which phone on-call responders about incidents.