About SchemaPilot

AI should answer from your data without being trusted with it

SchemaPilot exists because the teams who wanted a chatbot over their databases were the same teams who could not accept text-to-SQL. We built the layer in between: discovery, a reviewed semantic model, governed tools, drift review, and a Co-Pilot that can only run what you approved.

AI may propose, interpret, explain and select. It may not authorize, publish or execute ungoverned data access.
Principles

Four rules every feature has to pass

People authorize, AI proposes

The model may interpret a question, choose among approved tools and word an answer. Approving a tool, publishing it or executing a query is always a human action with a revision number behind it.

Read-only by construction

Connectors open read-only sessions, every identifier is quoted, every value is bound and results are capped. There is no code path that turns a prompt into executable SQL.

Reviewable artifacts

Every stage leaves something a person can read and compare: a catalog snapshot, a semantic model revision, a scored tool with its rationale, an audit entry with typed arguments.

Your keys, your tenant

LLM keys and database credentials are encrypted per workspace with the workspace id bound into the ciphertext. Nothing is shared across tenants and nothing is logged.

What we will not do

Guarantees you can hold us to

These are design decisions, not settings. They hold on every plan, during trials, and inside Co-Pilots you deploy into your own applications.

  • No demo data, seeded accounts or fabricated examples inside the product. What you see is your schema.
  • No raw SQL generated from natural language, in any mode, for any plan.
  • No row-level data, schema or credentials sent to a language model. Routing sees tool names and descriptions; wording sees at most twenty aggregate rows.
  • No silent changes to approved tools. Regeneration keeps approvals only where the definition is unchanged, and schema drift suspends exactly what it affects.
  • No lock-in on the model provider: the key is yours and can be rotated or removed at any time.
How we build

Governance first, then everything else

Deterministic where it matters. Schema analysis, tool scoring, plan validation and SQL compilation are deterministic and tested on synthetic catalogs. A language model is only invited where judgement helps: wording, naming, choosing among approved tools.

Honest about cost. Enrichment and enhancement use the workspace's active AI provider as background jobs you can watch and cancel. Results are stored so a page reload never repeats the model call.

Pluggable by design. PostgreSQL, Microsoft SQL Server and MySQL / MariaDB share one governance model; adding an engine means adding a connector and a dialect, never new rules.

Open about limits. Plan limits, trial length and supported engines come from the platform catalog you see on the pricing page, not from marketing copy. Read the Terms and the Privacy Policy.

Questions about how we handle your data?

We would rather you ask before you connect

Or email support directly.