Know what your SQL does before the warehouse runs it.

Static checks

A type checker for what a query means.

Semiring reads the compiled SQL of every dbt model and works out what each column is made of, which rows survive, and how many times each one can repeat. It answers from the code. Nothing is executed and no data leaves your warehouse.

Fan-out

Finds joins that can repeat rows, and the aggregates they inflate. Proven impossible when a key guarantees one match.

Nulls

Tracks NULL through outer joins, coalesce, case without else and empty aggregates, then checks it against your not_null tests.

Lineage

Column-level, through CTEs, windows, unions and flatten. Each path is graded proven, approximate or not traceable, so you know how far to trust it.

Keys and grain

Derives what makes a row unique from the SQL, and tells you which unique and not_null tests it can already back.

Incremental

Says what an incremental model needs to stay correct: its cursor, its lookback, the rows it must reread.

Drift

Compares the columns your code publishes with what the warehouse catalog says is deployed.

Semantic review in CI

A reviewer that reads the algebra, not the diff.

A forty-line SQL diff can be a harmless refactor or a silent change to revenue. Semiring compares the meaning of both versions, tells you which columns and rows can differ, and looks for a concrete row that proves it.

− from (select * from stg_payments
−       where paid) p
+ from paid p

− left join stg_shipments s
+ inner join stg_shipments s

−     o.customer_id,

Verdicts

Every changed model is equivalent, changed or broken, with the SQL fragment responsible and the evidence behind the call.

Counterexamples

For a changed model, Semiring synthesizes a small input, runs both versions locally and shows the rows that differ. Duplicates and NULLs included.

Blast radius

Lists the downstream columns a change can reach, and whether it touches their values, their rows, or both.

Smaller CI

Build and test the models a change can reach. Skip the hundreds it cannot.

Honest limits

When SQL is outside what it can analyze, the report says so and widens the answer. It never guesses quietly.

For coding agents

Give your agent something to check its SQL against.

Agents write SQL that looks right. Without a compiler that understands joins and NULLs, the first thing to tell them otherwise is production. Semiring gives them the same answers it gives you, as tool calls, with no warehouse credentials.

$ semiring impact fct_orders.revenue --json
{
  "column": "fct_orders.revenue",
  "downstream": { "models": 6, "columns": 8 },
  "affected": [
    { "column": "fct_revenue.revenue_usd",
      "effect": "value", "grade": "proven" },
    { "column": "rpt_finance_daily.revenue",
      "effect": "value+rows", "grade": "approximate" }
  ]
}

One call, one answer

explain, impact, criteria, diff: each a CLI command with a JSON form, usable from Claude Code, Codex or Cursor.

Graded evidence

Every answer says how sure it is, so an agent knows what it has established and what it still has to test.

Tests that matter

For any model, the input cases a test has to cover: match, no match, NULL key, multiple matches.

A safe place to run

The agent verifies its change on the local endpoint. It never needs a production role.

Desktop IDE

All of it, in one window.

A desktop editor for dbt projects where the checks, the lineage, the local warehouse and the agent share one index of your project. The demo at the top of this page is the idea.

Inline findings

Fan-out, NULL and broken-reference findings on the line that causes them, updated as you type.

Hover lineage

Point at a column to see where its value comes from and what reads it, across models.

Run locally

Build a model and its children on the bundled endpoint. Results in the editor, in seconds.

Review before push

See the semantic review your pull request will get, before you open it.

Early access

Tell us what to build first.

Semiring is in development and none of it is generally available yet. Leave an email and pick what you would use. The answers decide the order we ship in, and we will write when there is something for you to try.