Nobody's job is watching a dashboard. Write the rule in plain English, and an agent checks your data on your schedule — then speaks up, with receipts, only when something's real.
$50 credit · No card · Your first agent running in an afternoon
Runs on the same platform in production at WashU Medicine, Saint Louis University & dozens more · HIPAA-ready · Hosted or in your own cloud
01You write the skill
02It runs on your schedule
03It tells you what matters
01
You write the skill.
Brief it the way you'd brief a new analyst: what to look at, what counts as worth reporting, and what to do about it. Pick the tables from a picker, set the bounds, save.
“If the gap is under 5%, reply with exactly ‘Nothing to report.’ Otherwise: slice by segment and source, then explain the most likely cause — and say how confident you are.”
Not sure where to start? Describe what you want and Databasin drafts the skill for you — then you edit it like any other.
02
It runs on your schedule.
An agent is a task on the same canvas as your SQL, dbt and notebooks — drop it into a stage after your pipelines refresh and it reads data that's already current.
It gets read-only tools, a query budget, and a turn limit. Every step lands in the run log: which skill and version, which tools it was offered, every query it spent, how it finished.
03
It tells you what matters.
The finding goes to Slack, Teams, or your inbox — the number, the cause, the confidence, and a link to the query behind it. Not an alert that something moved. An explanation of why.
And on the mornings when nothing's wrong, it says so and gets out of your way. That's the whole point.
The unit of work
A skill is four things. None of them are code.
Most "AI agent" features are a prompt box an admin owns. A Databasin skill is a real object: you can read it before you run it, copy it under your own name, change it, share it, and roll it back. Because it's text, not code, the person who actually knows the business can write it.
Start from ours
Three skills, ready to schedule.
Every account ships with a curated library. Pick one, name your own copy, point it at your tables — you're running in minutes. When we improve the original, your copy tells you an update is available and you decide whether to take it.
Executive summary
Reads the fresh numbers and writes the narrative: one headline takeaway, three to five metrics with context, what moved and why, and two or three concrete next steps.
Monday, 7am, already in Slack
Data-quality sweep
Profiles a table for null rates, duplicates on your key columns, impossible values, and freshness — then rates it Good, Fair or Poor and ranks what's most likely to break your reporting.
Nightly, before anyone opens a dashboard
Anomaly & root cause
Confirms the metric really moved against a baseline, drills down by every dimension to find where the change concentrates, then ranks the likely causes with the evidence for each.
Whenever the number looks wrong
Then write your own — a renewals watch, a census check, a grant-burn monitor, a nightly reconciliation. If you can describe it, you can schedule it.
Where they run
Not a separate product. A task type.
Agents live inside the automations you already build, which is what makes them trustworthy: the agent runs after the pipeline that feeds it, on the same schedule, with the same retries, logs and versioned snapshots as every other task on the board.
Automation you can audit
Guardrails, not vibes.
The reason you can put this on a schedule and sleep is that the blast radius is small on purpose. A bad prompt produces a bad report — never a bad write.
Read-only tools, full stop
An agent gets three: get_schema, run_sql, fetch_data. A skill can narrow that list; nothing can widen it.
SELECT-only, verified
Every statement is checked before it executes. Anything that isn't a single SELECT is rejected by the runtime, not by the prompt.
Budgets you set
A turn limit and a query budget per run, plus a hard row cap on everything it reads back. Runaway loops aren't possible.
Governed data only
The agent queries the connection the task points at — your gold layer, not raw tables — and inherits that connection's permissions.
A receipt for every run
Skill and version, tools offered, each turn, each query, how it ended, and the report it produced — stored with the run.
Your model, your key
Claude or GPT, pinned to a connector you own. Self-installed? Your tenant, your LLM, billed to you with no markup.
Versioned like code
Every save snapshots the skill. Compare versions, pin one, or roll back — the automation keeps running the version it was given.
Metered, not mystery-priced
Runs bill like everything else: published per-minute rates plus the tokens the run actually spends. Stop the schedule, stop the meter.
Who gets to see it
Start private. Share when it works.
A skill you tune inside an automation belongs to that automation — nobody else is affected while you're still getting it right. When it earns its keep, widen it in one click.
Just this automation
Your copy, your tuning. Invisible in everyone else's picker, and editable right where you use it.
This project
The team working on this data can pick it up, run it, and fork their own variant.
Your whole organization
Every project sees it. Admins get a governance view of what people actually built — and can promote the good ones.
No admin ticket required. Anyone who can work in a project can write a skill — which is the point, because the person who knows what "worth reporting" means usually isn't the admin.
Straight answers
What agents don't do.
Worth knowing before you plan around them.
They don't write to your data.
Not a setting you can flip — the tools simply aren't there. Write-back and delivery are separate, explicitly configured steps that run after the agent finishes.
They don't act on their findings.
An agent produces a report. Opening the ticket, adjusting the forecast, emailing the customer — those stay with you, or with a task you wire up deliberately.
They don't run forever.
Each run is bounded by your turn and query limits. This is a scheduled analyst, not a daemon with an open-ended budget.
They don't replace your judgment.
Every claim comes with the query that produced it, precisely so you can check the ones that matter. Trust it, or verify it — the receipt is always there.
Ready when you are
Your newest analyst starts tomorrow at 7am.
Just your email — we'll build your workspace and send your sign-in link.
