Problem

On-call access: for the incident, not the shift

Also asked as: on-call access management, just-in-time access for on-call engineers, and how to give support temporary access on a weekend rotation.

The on-call schedule should decide who may ask for access. The incident should decide when they get it. The clock should decide when it ends.

On-call access is the access an engineer or support agent needs to act on a page or an escalation outside working hours. Done well, being on call gives you the right to request it without waiting for anyone, the grant lasts hours rather than the whole shift, and it ends on its own whether or not anyone remembers it at handover.

Why on-call is where standing access comes from

An on-call engineer has to be able to act at 3am, and nobody wants an outage to wait on a ticket. The cheapest way to guarantee that is to give everyone on the rotation the production admin role, permanently. A rotation eventually includes most of the team, so most of the team ends up holding it, all week, every week, including the weeks they are not on call.

Support rotations do the same thing with different systems: the weekend agent gets the customer-admin console “just in case”, and keeps it on Monday. Nobody decides to give a team standing access. The rotation decides for them, one incident at a time.

Three ways teams give on-call access

GrainWhat it looks likeHow it fails
Standing Everyone who is ever on the rotation holds the role, permanently The access exists when nobody is paged, which is nearly all the time. Any compromised laptop or token on the rotation can use it.
Per shift A tool reads the paging schedule, grants the role at handover and removes it at the next one On a weekly rotation that is 168 hours of standing access for whoever holds the pager, most of them quiet. It also makes your access policy only as correct as your schedule: every swap, override and “can you cover Tuesday” has to land in the paging tool before it lands in production.
Per incident The person on call requests access when they are paged; it lasts a few hours and removes itself Only if asking is slower than reading the alert. Then people keep standing access anyway.

Per shift is a real improvement on standing, and it is what several tools that read a paging schedule offer. Per incident is smaller still, and it does not need your paging tool to be right. It has the one failure mode, and the rest of this page is about avoiding it.

Split it into three questions

Most on-call access designs fail because they answer one question with a mechanism built for another. Taken apart, there are three:

  1. Who may ask? The rotation. Being on call, or being on the team that takes turns, is what makes someone eligible to see and request the on-call access. It is not, by itself, the access.
  2. Who approves? It depends on what is asked for, not on the hour. Read access to logs and dashboards needs nobody. Changing production needs a second person. Reading customer data needs someone outside the requester's own queue.
  3. How long? Hours, chosen before the access starts. Long enough for a bad incident, short enough that it is gone before the next shift has to think about it.

Kept apart like this, handover stops being an access event at all. Nothing has to be removed when the pager moves, because nothing was granted for the shift.

What has to be true at 3am

  • Read access is self-service. The on-call engineer asks for read access to logs and dashboards and has it within a minute, with no approver. The window and the record bound the risk; an approver at this hour adds only delay.
  • Write access has an approver who is awake. Name a group, not a person: the secondary on call, or the incident commanders. One named approver is one plane flight away from nobody.
  • The approving group is not the requesting group. If everyone on the rotation may approve on-call write access, the person on call can usually approve their own request, and the approval means nothing. Let the rotation ask and a different group decide.
  • There is a way through when nobody answers. A short wait, then the requester may proceed with a written reason, and everyone who should have answered is told immediately. The break-glass guide covers what that path has to get right.
  • The grant names the incident. Ask for the incident or ticket number as a required field on the request. A month later, “who had production write on the 14th, and why” should be one lookup, not an archaeology project in chat logs.
  • The incident channel sees it. A message when access is requested, granted and removed, posted where the responders already are, means a grant during an incident is never a surprise to the people working it.
  • Removal needs nobody. The person who was paged at 3am will not remember to give access back at 5am, and they should not have to.

Support on call

A support rotation looks different and has the same shape. The weekend agent needs, now and then, something the queue does not give them: an admin role in the customer-facing console, a group that opens the billing system, a read on a customer's records to answer an escalation. The same three questions apply:

  • Who may ask is the support rotation, or the support team as a whole.
  • Who approves depends on the data. A role that changes a customer's configuration wants the support lead. A read of a customer's personal data wants someone outside support, because a peer who shares the queue's deadline is the wrong check on it.
  • How long is the length of one escalation, not the weekend. Ask for the ticket number, and the record of who looked at which customer, when, and for which ticket writes itself.

Handover, swaps and the paging tool

Your schedule lives in a paging tool, and it changes constantly: swaps, overrides, someone covering an hour while another is on a train. Every one of those changes the answer to “who is on call”. If access is granted per shift from the schedule, every one also changes who holds production, and a stale override is a stale grant.

Per incident takes most of that weight off the schedule. The schedule only has to decide who may ask, and being off by an hour on that is a much smaller mistake. If your team keeps an on-call group in its directory, updated at handover by a script or by hand, that group is a good answer to who may ask. If it does not, the whole rotation is a fine one: being on the team that takes the pager is eligibility enough, because asking is not having.

It also means a change of paging tool is not a change of access policy. Teams on Opsgenie are living that now: Atlassian ended new sales in June 2025, and support ends on 5 April 2027. A team whose production access is wired to Opsgenie's schedules has two migrations to make, not one.

A grant that is running at handover can simply keep running. It ends within hours anyway, and the engineer who was paged can give it back early once the incident is closed.

When the directory itself is down

Every design above runs through your identity provider: sign-in, the on-call group, the groups that carry the access. On the night the identity provider is the incident, none of it works, and that is a different emergency with a different answer — emergency accounts that do not depend on the provider, held offline, alarmed when used. The break-glass guide separates the two, and a rotation needs a plan for both.

A starting configuration

Three packages cover most engineering and support rotations. Adjust the windows to your incidents; start short and lengthen only when an incident actually ran out of time.

PackageWho may askApprovalWindowAsked at request
oncall-prod-read The on-call group, or the whole rotation None — granted on request, recorded 4 hours Incident number
oncall-prod-write The on-call group The incident-commander or secondary group; break-glass after a short wait 2 to 4 hours Incident number, and what will change
support-escalation The support rotation The support lead group; someone outside support where customer data is read 2 hours Ticket number

How TemprBac fits

TemprBac handles the per-incident grain, and it is worth being exact about what that means.

  • It does not read PagerDuty, Opsgenie or any paging tool's schedule. If you want access granted at the start of a shift from the schedule, some of the tools in our just-in-time access comparison do that, and TemprBac does not.
  • Who may ask, and who approves, can each be a group in Entra ID, Okta, Google Workspace, GitHub, AWS IAM or Active Directory. Members are read from the directory and cached for up to 15 minutes, so a handover takes up to that long to show. A group that cannot be read admits nobody: it fails closed.
  • An approver can decide their own request in the ordinary path, which is why the approving group above is not the on-call group.
  • Every grant lasts one to eight hours, set per package, and is removed on its own when the window closes. There is no setting for “never remove”. Leaving the on-call group does not cut a grant short; the engineer, or an admin, can remove it early.
  • Break-glass is off by default. Turned on, it lets a requester approve their own request once it has waited 5 to 60 minutes, with a written reason, and emails the package's approvers and the org's admins at once.
  • A package can ask its own questions, such as an incident number, and post each request, grant and removal to a Slack channel or a Teams workflow.

Access for the incident

The rotation may ask, the right person approves, and the grant ends on its own — before the next handover has to think about it.

Join the waitlist