# Dry — build a full app from one JSON schema

> Dry is the backend: database, permissions, hosting, and rendering already exist.
> Describe an app as JSON and deploy it with ONE HTTP call. Do NOT build a server,
> database, API, or auth of your own.

## If you are an AI agent, do this
There is one schema. **If you can write files, author it as a `.dry/` folder on disk
FIRST, then deploy from that folder.** The `.dry/` files are the deliverable the user
keeps — they can edit and re-deploy them. Deploying an in-memory bundle without ever
writing the files leaves the user nothing on disk; only do that if you have no
filesystem at all.

- Shell / coding agent (can write files): read the docs below, WRITE the `.dry/`
  folder to disk, then deploy that folder (assemble its files into the bundle and
  POST — or run `clients/deploy.mjs .dry` / `docs/tools/dry_deploy.py .dry/`).
- File-writing, no shell: create the `.dry/` files from the schema below, then POST
  them as one bundle.
- Chat / HTTP only, NO filesystem: only then skip files — build the bundle and POST it.

**`import-app` is the ONLY deploy path here.** Do NOT use `create_app_space` /
`createAppSpace`, nor per-item `create_item` / `createItem` / `import_items`. Those
generate the types and pages SERVER-side from a prompt and leave the user NO local
`.dry/` files — that defeats the entire point of this kit. YOU author the types, pages,
and seed data as `.dry/` files, then deploy them with one `import-app` call.

Deploy (all tiers):
  POST https://unsupervisedlearning.dry.ai/api/dbcrud
  { "op": "import-app", "bundle": { "app": {...}, "types": [...], "pages": [...], "data": [...], "permissions": {...} } }

## Auth (one time, no browser needed) — all on this same endpoint
Every deploy call sends the header:  Authorization: Bearer <token>
Get a token with two calls to the SAME /api/dbcrud endpoint — a 6-digit email code
is the only manual step:
  1. POST https://unsupervisedlearning.dry.ai/api/dbcrud
     { "op": "register", "email": "you@example.com" }
     -> emails a 6-digit code; returns { "userId": "..." }
  2. POST https://unsupervisedlearning.dry.ai/api/dbcrud
     { "op": "verify", "code": "123456", "userId": "<from step 1>", "email": "you@example.com" }
     -> returns { "mcpToken": "..." }.  Use that mcpToken as your Bearer token.
(Prefer an API access key? Once authenticated: POST https://unsupervisedlearning.dry.ai/api/dbcrud {"op":"generate-key"}
returns one. The same token also works on /api/custom-gpt and the MCP endpoints. Already
signed into Dry in a browser? GET https://unsupervisedlearning.dry.ai/geak mints a key too.)

## Docs
- Instructions (canonical):        https://unsupervisedlearning.dry.ai/agent-kit
- Format reference (types/widgets/helpers): https://unsupervisedlearning.dry.ai/agent-kit/dry-format.md
- JSON Schema (validate first):    https://unsupervisedlearning.dry.ai/agent-kit/schema/bundle.schema.json
  bundle.schema.json references 5 sub-schemas at the same path — fetch+register all:
  app.schema.json, type.schema.json, page.schema.json, data.schema.json, permissions.schema.json
  Schemas are JSON Schema draft 2020-12 — use a 2020-12 validator (e.g. Ajv2020), not draft-07.

## Deploy loop
1. Assemble the bundle from your `.dry/` files (or build it directly only if you
   have no filesystem); validate it against the JSON Schema.
2. POST it -> { space:{id}, pages:[{title,status,id,viewPath}] }.
   A page with "html" is stored verbatim (ready now); a "prompt"-only page is
   generated in the background.
3. Poll until ready:  POST https://unsupervisedlearning.dry.ai/api/dbcrud { "op":"app-status", "space":"<id>" }
   until data.generating == 0.
Re-POSTing the same bundle is idempotent (upsert by name). On a bad bundle you get
HTTP 400 + errors:[{path,message}] — fix that field and resend.

## Files (uploads)
A picture/file field value may be an uploaded file:
  { "filename":"logo.png", "contentType":"image/png", "data":"<base64 bytes>" }

## AI-computed fields
Mark a field `"aiGenerated": true` (+ a `"prompt"`) and Dry fills its value
automatically on import, derived from the item's other fields (e.g. estimate a
meal's calories). Omit that field in your seed data and it generates. See
https://unsupervisedlearning.dry.ai/agent-kit/dry-format.md -> "AI-computed fields".