On-call schedules and paging

Rotations, overrides and escalation policies that page whoever is on call when an alert fires, by email, Slack, webhook, text or phone, until someone acknowledges.

On-call decides who answers when an alert fires, and keeps reaching people until someone does. A schedule says whose turn it is. An escalation policy says who is paged first, how long to wait, and who comes next. An alert rule that pages a policy opens an incident for each alert, and the incident pages step by step until someone acknowledges it or resolves it. Each member chooses how they are reached: email, Slack, a signed webhook, or a text message or phone call.

Schedules

Open On-call in the dashboard and add a schedule. Schedules belong to the workspace, so one schedule can cover every project in it. A workspace can have up to 20 schedules. Each schedule has a name and an IANA time zone, such as Europe/Paris or America/New_York, UTC by default. Every time and date in the schedule is read on that time zone's wall clock.

Layers and rotations

A schedule has up to 5 layers. Each layer is a rotation: its members, up to 30, take turns in the order you list them. The same person may appear more than once.

  • Daily: the turn passes to the next member every day at the handoff time.
  • Weekly: the turn passes once a week, at the handoff time on the day you choose.
  • Start date: the first turn starts at the first handoff on or after this date. Turns follow in order from there.

A handoff at 09:00 stays at 09:00 on the schedule's clock when daylight saving starts or ends, so the turn on those days is an hour shorter or longer.

A layer can be limited to some days of the week and some hours, such as weekdays from 09:00 to 17:00. An end time at or before the start time runs past midnight into the next day, and equal times cover the whole day. Outside its hours, a layer covers nobody.

Layers stack. A higher layer covers the ones below it wherever it has someone, and a lower layer shows through wherever it does not. A common setup is a weekly rotation for nights and weekends at the bottom, with a weekday daytime rotation above it.

Overrides

An override puts one member on call for a time range, over every layer: someone covers a shift while a teammate is away. An override can end up to 90 days ahead and last up to 90 days. A schedule keeps up to 200 overrides that have not ended. Where overrides overlap, the one made last wins. Someone who left the workspace is passed over: the override before theirs, or the layers below, cover their time.

Who is on call

Each schedule shows who is on call now, whether a layer or an override put them there, and until when. Its calendar shows the turns to come, 14 days by default and up to 90, with every layer and override already combined, and gaps where nobody is on call.

Escalation policies

An escalation policy is a list of steps, in order. A workspace can have up to 20 policies, each with up to 10 steps.

  • Who: each step pages up to 10 targets. A target is a schedule, which pages whoever is on call on it at that moment, or a named member.
  • Wait: how many minutes the step waits before the next one, 1 to 1,440, 15 by default. After the last step, it is the wait before the policy starts again.
  • Repeat: how many more times the whole policy runs after its last step, 0 to 9.

A step that reaches nobody, such as a schedule with a gap at that moment, moves straight on to the next step without waiting. Once the last run is over, no more pages go out, and the incident stays open until someone acknowledges or resolves it, or its alert recovers.

Alerts and incidents

In Settings, Alerts, choose an escalation policy for an alert rule. The rule still sends its emails and Slack posts as before, and each alert it opens now also opens an incident. Incidents are numbered 1, 2, 3 and so on within the workspace, and carry the alert's words, such as Check down: api health.

An incident is in one of these states:

  • triggered (Triggered): paging, step by step, as its policy says.
  • acknowledged (Acknowledged): someone is on it. Escalation stops, and nobody else is paged.
  • resolved (Resolved): over. Nothing more is paged.
  • An incident resolves on its own when its alert recovers, such as an uptime check passing again or a cron job that ran.
  • New error and regression alerts are over when they are sent, so nothing recovers on its own. A person resolves their incidents.
  • An alert inside its rule's cooldown sends no email, but still opens its incident and pages: the cooldown never holds back a page.
  • Turning a rule off or deleting it, or deleting its project, closes its open incidents.

Each incident keeps a log of every step it ran, who each step reached, and who acknowledged or resolved it and how. Resolved incidents are kept 400 days from when they opened.

Acknowledge and resolve

There are three ways to answer an incident:

  • On the On-call page in the dashboard, or the incident's own page.
  • With the Little Friend CLI, which also shows who is on call now and the calendar to come.
  • With the one-click links in each page.
On-call from the CLI
littlefriend oncall now
littlefriend oncall calendar Primary --days 30
littlefriend incidents list --status open
littlefriend incidents show 42
littlefriend incidents ack 42
littlefriend incidents resolve 42

An incident is named by its number or its id, and a schedule by its name or its id. The calendar shows 14 days unless you pass --days, up to 90.

A link in a page does one thing, acknowledge or resolve, for one incident, and works for 24 hours. It opens a page that shows the incident and a button, and only pressing the button acts. A mail scanner or a link preview that opens the link changes nothing.

AI assistants connected over MCP can read who is on call with the get_on_call tool. It only reads: acknowledging and resolving take the dashboard, the CLI or a link.

How you are reached

Each member chooses their own channels on the On-call page, and every page goes to all of them. Up to 10 channels per member, besides their account email:

  • Email: the member's account email, always, while it is verified.
  • Slack: a Slack incoming webhook, set up as for alert posts, with an optional Slack member id, such as U012AB3CD, to mention in each page.
  • Webhook: a signed POST to your own URL. See below.
  • Text message and phone call: when the server has a phone provider set up. See below.

Send a test to a channel to check it. A channel that stops working is marked broken, shows why, and gets nothing until a test goes through.

Webhooks

A webhook channel gets a JSON POST for each page and notice. Its type is one of incident.triggered, incident.acknowledged, incident.resolved or test. A page carries the incident's link and its acknowledge and resolve links. A test carries no incident.

A page
{
  "type": "incident.triggered",
  "deliveryId": "d8Jq3VmT0pXc5LwN2rKa",
  "sentAt": "2026-10-26T14:30:05.000Z",
  "incident": {
    "id": "inc_4fT9kQ2mXr7Lp1WzBa8c",
    "number": 42,
    "title": "Check down: api health",
    "kind": "check_down",
    "status": "triggered",
    "projectId": "prj_7Hc2",
    "projectName": "Example",
    "openedAt": "2026-10-26T14:30:00.000Z",
    "url": "https://app.littlefriend.io/p/prj_7Hc2/oncall/incidents/inc_4fT9kQ2mXr7Lp1WzBa8c",
    "acknowledgeUrl": "https://app.littlefriend.io/oncall/link#token=lfo_Ak3mQ9xLp2Nw7aZcRt5mVb8Yd1Kf_e4H",
    "resolveUrl": "https://app.littlefriend.io/oncall/link#token=lfo_Rz6tW1cYe3kdP8vN0bJh2Lq7Ms9Xg-a5"
  }
}

Every post is signed. x-lf-signature is v1= and the hex HMAC-SHA256 of <x-lf-timestamp>.<body>: the x-lf-timestamp header, a dot and the raw body, keyed with the channel's signing secret. x-lf-timestamp is the time the post was sent, in Unix seconds: refuse a post whose timestamp is more than 5 minutes off your clock, so a captured post cannot be sent again later. The secret is shown once, when you add the channel: keep it in your app's secrets.

A post can arrive more than once, for example when a try times out after your app got it. x-lf-delivery and the body's deliveryId are the same on every try of one post, so drop a repeat by them.

Check the signature in Node.js
import { createHmac, timingSafeEqual } from 'node:crypto';

// rawBody is the request body exactly as received, before JSON parsing
export function fromLittleFriend(rawBody: string, headers: Headers, secret: string): boolean {
  const timestamp = headers.get('x-lf-timestamp') ?? '';
  const signature = headers.get('x-lf-signature') ?? '';
  // Refuse posts more than 5 minutes off, so a captured one cannot be sent again later
  if (!(Math.abs(Date.now() / 1000 - Number(timestamp)) <= 300)) return false;
  const expected = 'v1=' + createHmac('sha256', secret).update(`${timestamp}.${rawBody}`).digest('hex');
  return signature.length === expected.length && timingSafeEqual(Buffer.from(signature), Buffer.from(expected));
}

Text messages and phone calls

Text messages and calls work when the server has a phone provider set up. Little Friend's own hosted service sends them with AWS End User Messaging. Enter a number in international form, such as +1 415 555 0123, and confirm it with the 6 digit code sent to it. A code works for 10 minutes and 5 tries, and a new one can be sent 60 seconds after the last. Adding a number needs a verified email address and a workspace, and the hosted service texts and calls numbers in the US and Canada only.

The number is kept sealed and shown masked, such as +1 ••• ••• 0123. It is used only for paging: never in analytics data, and never written to a log. One member gets at most 20 texts and calls an hour, across every workspace.

What is sent, and when

  • Pages go to every channel of each person a step reaches.
  • Acknowledged and resolved notices go to email, Slack and webhooks only, and only to the people already paged for that incident. They never text or call.
  • Each page or notice gets up to 4 tries. A page not sent within 30 minutes of its step is dropped, since the next step is already on its way. A notice not sent within 6 hours is dropped.

Each incident's page lists every page and notice, who it went to, and whether it was sent.

Privacy

A page carries the alert's words, the project's name and links to the dashboard. It never carries a visitor, a session, a URL with a query or an email address. Phone numbers stay out of analytics data and logs, as above.