Replay-minute
A replay-minute is one minute of sandbox time an agent spends executing against a Mirrors environment, either replaying a recorded session or running an eval. It is the unit Mirrors bills. Building an environment, browsing it, and reading results cost nothing; the meter runs only while the agent is running.
Last updated 2 August 2026
What starts and stops the meter
The meter starts when a sandbox begins executing your agent against an environment and stops when that run finishes. It measures wall-clock time inside the sandbox: not model tokens, not the number of tool calls, and not the time you spend reading the result afterwards.
| Action | Metered |
|---|---|
| Building an environment from traces, code, or docs | No |
| Browsing an environment, its schema, or its seed data | No |
| Replaying a recorded session against a change | Yes, while the run executes |
| Running an eval suite against an environment | Yes, while the run executes |
| Reading a replay diff, a report, or the dashboard | No |
Two runs executing at once consume two minutes per minute of wall-clock time. The meter follows the sandbox, not the person waiting on it.
Why a minute, and not a run or a seat
A run is not a stable unit. One replay is a three-step lookup; the next is a forty-step workflow across six systems. Priced per run, the cheap case subsidizes the expensive one and teams start rationing the exact thing the product exists for.
Seats are worse. An agent test suite is run by CI, not by people, so the number of humans on the team says nothing about the load. Sandbox time is what costs us money to provide, so it is what we charge for, and it moves with CI volume, which is a number a team can predict and control.
What it costs
Free is $0/month and includes 60 replay-minutes each month; after that, usage is $0.20 per replay-minute, billed monthly. Enterprise is custom: volume pricing with SSO, role controls, a security review, and on-premises or BYOC options. Those are the only two plans.
How it differs from the units next to it
| Unit | What it measures | Who bills in it |
|---|---|---|
| Replay-minute | Wall-clock sandbox time while an agent executes against an environment | Mirrors |
| Token | Text volume into and out of a model | Model providers |
| Trace or span | One recorded execution, or one step inside it | Observability and eval platforms |
| Seat | A person with access | Most developer tooling |
Model usage during a replay is billed by the model provider, under whatever key the run uses. A replay-minute pays for the sandbox around the model, not for the model.
Estimating what you will use
Measure it rather than model it: replay one representative session and read the elapsed time off the run. As a rough shape, a short support-agent session finishes well inside a minute, while a long multi-tool workflow waiting on slow model calls is the case that runs into several. What turns that into a monthly number is CI: a gate that replays a suite on every pull request multiplies one session by your merge rate.
Related terms
- Agent regression test: A replay of a recorded session against a change, checking that behavior a team already relies on still holds.
- Agent staging environment: A non-production copy of the systems an agent calls, so its actions land on fabricated data.