This file is the app-specific design brief: scope, information architecture, and the screens. It does not restate visual tokens — colors, type, spacing, the five "Clawnify signature" moves, and dual human/agent mode all come from the Clawnify Apps design system, which every build inherits. Follow that for anything visual; follow this for what goes where.
OpenKitchen owns exactly one thing nothing else on the platform models: the
kitchen cost graph — ingredient ↔ supplier price ↔ recipe/prep ↔ food cost ↔ inventory ↔ production ↔ supplier invoice — plus a documents lifecycle for
licenses/permits/HACCP.
It borrows the AI-native pieces instead of rebuilding them:
- Invoice OCR → the multimodal agent reads the photo and POSTs lines. No OCR service in this app.
- Forecasting / P&L narrative / variance analysis → the agent over the JSON
API (
OPENROUTER_API_KEYfor prose; deterministic heuristics otherwise — same pattern asopen-ads-report). - Statutory books, VAT, e-invoicing →
open-books. OpenKitchen surfaces food cost and margin; it does not grow a second ledger.
The app therefore does the deterministic math (cost roll-up, food-cost %, variance, reorder points, cost drift) and exposes it; the agent does the judgement. This split — record vs action — is the product story: your kitchen's back office, run by an employee you text, not another dashboard you log into.
React 19 + Hono + @clawnify/db on Cloudflare Workers + D1, Tailwind v4, driven by
the clawnify CLI — matching open-books. Server: src/server/{index.ts, db.ts, routes.ts, domain/*}, schema in src/server/schema.sql. Uploads (invoice & document
scans) go to R2 via src/server/uploads.ts (storage: true). UUID string PKs,
generated server-side with crypto.randomUUID(); money in integer cents.
Grouped, each group titled with a muted eyebrow:
- Overview — Dashboard
- Menu — Recipes, Ingredients, Suppliers
- Operations — Inventory, Production, Invoices
- Compliance — Documents
The one coral primary action is per-screen and contextual (New recipe / New ingredient / Log count / New invoice / New document).
Every screen is zoned cards opened by eyebrows; numbers are tabular with footer aggregates; enumerable values (category, unit, allergen, status) are chips, and attention states (over-target, expiring, unreconciled) are tinted badges.
- Dashboard — a KPI row (
GET /api/state): weighted food-cost %, stock value, open invoices, expiring documents. Then zones: Dishes over target (worst food-cost % first), Cost drift (ingredients whose price moved most, fromsupplier_price_history), Reorder now (at/under par), Expiring soon (documents). Empty state is borderless with the one next action. - Recipes — list with rolled-up cost / yield unit, price, margin, food-cost %
badge (green under target, amber/red over). Detail page: a Yield zone, a
Bill of materials zone (ingredients + nested preps, each line showing qty·unit
and its cost contribution, footer = total plate cost), and a Cost history zone.
A
preplooks identical but has no menu price. - Ingredients — table: name, category chip, base unit, current cost/unit, supplier, par level, allergen chips. Row → cost-history sparkline. Cost is read-mostly here; it changes via invoice reconciliation, not hand-editing.
- Suppliers — directory; detail shows that supplier's ingredients and recent invoices.
- Inventory — per ingredient: latest count, theoretical on-hand, par, and a
below-par badge. Primary action Log count (a
Dialog, not a prompt). - Production — log of runs (recipe · qty · date · source chip). This is what drives theoretical usage; central kitchens (e.g. eatdis) live here, POS imports land here too.
- Invoices — list with supplier, number, date, total, and a status badge
(
pending/reconciled/paid); a source chip marksagent-ocr. Detail shows scanned image alongside editable lines, each matchable to an ingredient; the coral action is Reconcile (updates costs + price history). Most invoices arrive already drafted by the agent — this screen is the human's review surface. - Documents — licenses/permits/leases/HACCP with issuer, reference, expiry, and
a lifecycle badge (
valid/expiring/expired). The agent nudges; this is where a human files the renewal.
Per the platform system: detect ?agent/mode=agent, set data-agent, bump targets
to 40–44px, every action a real <button aria-label>, semantic tables, no
alert/confirm/prompt/localStorage. Both the human (dashboard iframe) and the agent
(computer-use browser) drive these same screens.
With an empty database the app seeds sample data (a small menu, a few suppliers, one expiring HACCP certificate) so the dashboard looks alive inside the Clawnify iframe before the owner has entered anything — same convention as the other templates.