Spec Assist

Authoring a connector means knowing the spec's shape — the recordsPath, the idPath, a JSONata transform, the typed output columns. Spec Assist removes the blank page: paste a sample response, and it drafts those for you as a starting point you review and edit in the builder.

How it works

On New connector, choose Draft from a sample:

1. Enter the endpoint (method + path) and paste one JSON response body.
2. Spec Assist drafts a resource — recordsPath, idPath, updatedAtPath (if it finds one), a first-pass transform mapping the record's scalar fields to UPPER_SNAKE_CASE columns (always including externalId), and typed output.fields.
3. The draft is assembled into a spec and opened in the builder, pre-filled — you review, fix the auth/base URL, adjust the mapping, and publish.

The result is a suggestion, never auto-published — the same validation gates it as any hand-written spec, and you own the final edit.

Guardrails

• Secrets never leave your browser unredacted. Before the sample is sent to the model, credential-shaped fields (anything matching token / secret / password / apikey / authorization / …) are stripped — recursively, including nested objects.
• The sample is bounded. Large payloads are truncated to cap token cost.
• It only maps what's present. The draft never invents fields that aren't in the sample; it picks a real, unique idPath and guards nullable JSONata functions.
• Human-in-the-loop. The builder is the review surface — nothing is saved until you publish.

What it drafts (and what it doesn't)

Drafts, today: recordsPath, idPath, updatedAtPath, the transform expression, and typed output.fields for one resource.

You still choose: auth strategy and credentials, the pagination strategy, sync mode and change filtering, and any enrichment — Spec Assist gets you past extraction and mapping, which is the tedious part.