Graphdraft catches schema mistakes. Touchstone catches the gaps underneath them: give it one model and the questions your business will actually ask, and it tells you — question by question — which ones the model can answer, what is missing, and whether the fix is a property you forgot or an entity you never modelled.
No account, no install, no database connection. It reads the schema alone, so a model can be checked before any of it is built.
Public beta. Shared fair-use limits apply. Your work is kept in your own browser so a refresh does not lose it — never on the server.
Your model is not stored on the server. It is sent to the checker and on to Google's Gemini API to be analysed, then discarded — there is no database, and the logs record sizes and verdicts, never your schema or your questions.
A model can be perfectly well-formed and still be unable to answer a single thing the business cares about. These tools check different halves of "ready".
Dangling relationships, duplicate names, properties with no type, endpoints that point at nothing. Structural mistakes, caught while you draw.
Answered by the modelling tool, as you work.
"What is the total outstanding invoice amount per organisation?" — can this model answer that, today, with the labels and relationships it actually declares? A valid model fails this daily.
Answered against your real question catalogue, before production.
Every verdict is computed from your declared model, not asserted by a language model. The model proposes what the question needs; the checker decides what exists.
ANSWERABLE as it stands · ADDITIVE a property or relationship on labels you already have · STRUCTURAL the model has no home for it.
The middle one is the useful one: it separates "you are a property short" from "this model cannot express the question".
Start, traverse, filter, aggregate, return — in order, with every step that depends on something missing marked as such. Unanswerable questions still show their plan, with the blocking step flagged.
The one relationship that unblocks five questions, at the top. Additive fixes are listed separately from structural gaps, because they are different sizes of decision.
Graphdraft models load directly. So does whatever else your graph lives in — the file is converted to one declared schema, shown to you, and editable before anything is checked.
Generated queries are checked against your engine — Neptune, Neo4j, or portable openCypher — so you are never handed a query your database would reject.
Bring your own schema, or load one from the gallery and ask it something hard.