NerveStax — AI agents on the data stack you already run
A model change arrives as a tested pull request with its downstream list attached. An alert reaches you already diagnosed, or not at all. On your dbt, your Airflow, your warehouse, your repo. Nothing merges without a person.
Private beta · hosted or self-hosted · your own model keys
What lands on your side
Guaranteed by design
They do not depend on tuning, or on how long you have used it. They are how it is built.
If triage crashes, times out or is unsure, it pages anyway. Every escalation is written down before it is delivered.
Agents cannot merge and cannot push to your branches. A person approves before a pull request opens or anything runs in your orchestrator.
Built in an isolated sandbox against a development target, with the models and schedules it touches listed before anyone reviews it.
Everything it produces is ordinary code in your repository. Your pipelines run without it.
Three people, one week
01 · The engineer who asked for a model
The model, its tests, the run results and the list of what it touches downstream, opened in your repository. You review it the way you review any other pull request, and merge it or send it back.
02 · The person on call
The upstream that landed late, the retry that already succeeded, the echoes of a single failure: each is closed with a written reason you can read in the morning. The one real incident reaches your pager with the cause found and the next step written.
03 · The lead who gets asked what will break
Every change lists the models, tests and schedules that depend on it before anyone reviews it. Models, runs, freshness and drift sit in one record, so nobody explains the platform twice.
How it ships
Agents do not write to production. Your CI and review rules still apply.
A question from your team, or a failed run.
Built in an isolated copy, against a development target.
Your tests run before anyone reviews it.
Opened in your repository, with results and reasoning.
A person on your team merges it, or sends it back.
On your stack
dbt and Airflow, on Snowflake, Trino, DuckDB and MotherDuck, Postgres or ClickHouse. GitHub or GitLab, including self-hosted. PagerDuty, Opsgenie or Splunk On-Call for paging. Hosted on our cloud, or self-hosted on your Kubernetes, with your own model keys. Nothing to migrate.
All integrations →Questions
For a change: a pull request in your own repository with the code, its tests, the run results against a development target, the models and schedules it touches, and the reasoning. For an alert: a verdict, either closed with a written reason or escalated to your pager with the cause and the next step.
Read, build and test in its own sandbox. It cannot merge and cannot push to your branches. A person approves before it opens a pull request, triggers an Airflow run, runs a backfill or runs an ingestion. On-call triage is read-only.
Every verdict and every pull request carries its full transcript, each log it read and each check it ran, so a wrong diagnosis is visible in seconds rather than found in a retro. If triage is uncertain or fails, it escalates anyway. A wrong change is a pull request you decline.
dbt and Airflow, on Snowflake, Trino, DuckDB/MotherDuck, Postgres or ClickHouse, with GitHub or GitLab. Hosted on our cloud or self-hosted on your Kubernetes, with your own model keys. The private beta is free; beta teams get founding-customer pricing at launch. See pricing.
Private beta
Beta workspaces open in small batches. Start with the pipeline nobody wants to be on call for, or the model you have been meaning to build. Tell us what you run and we reply within a working day.