Seven questions a data lead should ask before installing On-Call, with the limits included. Each answer is tagged like the rest of the site.
Live
Production changes only through your git and your own deploy, after a named person signs. The agents’ working copy has no git remote: only On-Call writes to your repo, and only after a signature. For your warehouse, the boundary is the key you give the agents (next question).
Only after a named person signs does On-Call open a pull request with the fix. Your branch protection, required checks and reviewers apply, and the merge is yours.
Live
A warehouse key for the agents. Scope it to a dev dataset and cap its bytes: that scope is the boundary, and we’ll say so before you install.
The roles that build run dbt against a dev connection to your warehouse.
The agents work without asking permission for each command, so they can read whatever that key can read. What contains them is the key’s scope and the byte cap, not their good manners.
On-Call also needs a git repository it can open pull requests on, and a Claude subscription for the agents.
Live
No. Signing is a person, with their own account.
Signing happens on The morning, on the machine where On-Call runs. That screen only listens on the machine itself.
The owner of the model’s dbt group signs. If they can’t, your default signer does. If neither can, the case is escalated with the reason written down.
A versioned map decides what the crew may do without a person. If the map is missing, a person decides.
Live
A second agent, in a separate copy of your project, tries to refute it before you see it. The engine enforces that step: a fix does not reach you without it.
In build We also grade the crew on failures we cause on purpose: each broken case runs several times, and code scores whether the agent looked at the rows and whether its fix turns the test green.
Live
Monthly caps on estimated AI spend, per company and per agent, plus step, time and per-case limits.
When a cap is reached, work pauses. Each run is capped in steps and minutes; each case at 50 minutes, and after that agents stop and the case waits for a person. Warehouse spend is capped by your own dbt profile.
Live
The calls to the AI model, including what agents read from your warehouse; the queries to your warehouse; the branch and pull request to your git host; and the downloads on first install. The containers are not network-isolated: if you need that, run On-Call behind your firewall.
Sample rows from tests on columns or models tagged as personal data are kept out of the case. That does not stop an agent from reading other rows its key allows, and what it reads goes to the model.
Live
Every case keeps its evidence in your installation, and every landed commit records the case and the signer’s account.
Those records are three plain lines at the end of the commit message: the case, the source commit and the signer’s account identifier. They are plain text: a trace, not a digital signature.