Private Preview - features may change without notice
GramSpec

Everyone needs data foundations.
Almost nobody says what one is.

A foundation is the set of decisions about what your data means — what one row is, which things connect, what is allowed to be missing. Moving data can be computed. Meaning has to be decided, and the failures live at that edge.

Build one now → Read the spec → no signup · nothing to install
Live Describe a business in one sentence. Watch the decisions get made.

The notation is real — this is the spec, drawn. The practice is a sample, so nobody has to hand over data to see it work.

The method

Facts you can read out loud.

Every line the guide drew is one sentence with its constraints attached. That is the whole trick: if the sentence is wrong, the model is wrong, and anyone who knows the business can see it without reading a query.

Each Order is placed by exactly one Customer. one fact, end to end
What it carries
A relationship, a cardinality, and a participation rule — stated in one sentence, held even where the schema is silent about it.
What it determines
An inner join, the key it joins on, and grouping that cannot double-count. Computed from the constraints, not guessed.
Who can check it
Anyone who knows the business. No SQL required to spot that it is wrong.
Why it holds
An order exists or it does not. Business already runs on a closed world, and that assumption is what makes the numbers worth having.

Bring your own agent

We built four. You can build your own.

The specification is the backbone; our tools are one expression of it. Point the agent your team already runs at the same graph and it asks the questions instead.

$ claude mcp add --transport http gramspec https://gramspec.com/mcp
list_projectsget_graphgraph_overviewsearch_graphdescribe_entitygenerate_sqlvalidate_sqlrun_query

Eight tools arrive. Running a query passes the same checks as our own investigator, so the agent finds out what is missing rather than what is plausible. You bring the model and the database; model choice, spend and data residency stay yours.

Read the API reference →

The star schema, built

Starmaker

Settles what you measure with the people who run the process, then builds the star, loads it, and tests it against your rows.

The modeling canvas

Grammar

Refine the model in sentences on a visual canvas. The documentation regenerates from the model every time you open it.

The number you asked for

Vision

Ask at the moment you think of it, and the answer comes back drawn — filtered at the source, not in the browser.

The investigator

Reason

Give it a “why” question without choosing a graph first. It checks whether the premise even holds before it answers.


How it works

Each decision becomes a rule every query obeys.

Two decades of investment went into pipelines, storage and compute. The layer nobody built is the one that says what any of it means. The DDL never reaches the model — the rules do.

01

The model

A tool-using LLM, harness, or your own app. It never sees your schema.

02

The graph

An open spec for the rules: binary facts, cardinality, mandatoriness, role names.

03

The data

Any relational database, read-only, exactly where it already lives.

Who owns the meaning

The business takes the meaning back.

Whether a field may be empty is a business decision, not a database detail. These five keep the record of who decided what.

01

Accountability

Name a steward on an entity, or on the connection so everything beneath inherits it.

02

History

Versions seal at the rhythm of the work, a sitting at a time. “Customer is now Client.”

03

The record

One file: sorted, stripped of internal ids, identical to the byte from app or API.

04

Access

Graphs and dashboards are shared per person, by grant — never by handing over a credential.

05

Drift

A connected graph keeps diffing itself against the live schema and tells you what moved.


Trust

Your data stays where it is.

Every feature that touches the database was built security-first. Nothing is copied out, and nothing is written back.

Impossible, not improbable

Structural hallucination is unrepresentable.

A join that does not exist in the graph cannot be expressed, so it cannot be generated. This is not a low error rate that we measure and hope about — it is a shape the system has no way to draw.

Read-only

Validated server-side on every request. Writes blocked at the application layer.

Credentials

Encrypted at rest. Never sent to the browser, never placed in a prompt.

Two-factor

TOTP on every account, with any authenticator app.

Audit

Every execution logged with user, timestamp, duration and result.


Open standard

Read the spec. Build the foundation.

The graph is one JSON document you own. Take it with you, diff it, keep it in your repo. There is no gate on correcting it: a fix lands when the person who spotted it makes it.

Research

Lexical Independence of Schema Semantics — the constraint shape carries the meaning; predicate verbs in the graph are optional.

Read the paper (PDF) →

0

Hallucinated structure

Closed

By construction

8

Tools over MCP

Any

Relational database