Agents, with taste.

AI agents that make ads, landing pages and emails, guide shoppers in a store,
and run errands on your Mac, all starting from what people actually love.

---

Slice is five AI agents on one taste graph. Each agent starts with a read of its audience from Qloo's cultural taste graph (the artists, films, books, brands and places people love, and what those say about them) and calls Qloo as a tool while it works. That lets it personalize for a city, an audience or a single customer, even on a first visit with no history, and every result shows the evidence behind it.

Agent	What it makes
Media	Posters, ad creatives and short videos for a taste audience
Pages	Landing and funnel pages, one per audience or per lead
Email	Campaigns with a version for each taste group or person
Concierge	A shopping assistant that recommends from a store's own catalog
Slice for Mac	A voice agent in the notch that works across your apps

Try it

Open slice-agents.vercel.app and choose Continue as guest. You get a workspace that already holds two demo brands for every service, no sign-up needed.

Media. Designing a Half Moon Coffee poster for NYC film lovers.	Pages. Building a Tágide Festival page for visitors from Berlin.
Email. A Covers campaign, with restaurants ranked by taste.	Concierge. The widget on Marlow, with picks tuned to the shopper.

Products

Each service stands on its own: a business uses the ones it needs, and each has its own clients and data.

Media

An infinite canvas where a brand makes posters, ad creatives and short videos for one audience at a time. The agent reads the audience's taste, asks a couple of quick questions about the offer, then generates with Nano Banana (Gemini image models) and Veo 3.1. Every creative comes with a Why report built from Qloo signals and a simulated pre-test panel.

Demo brands: Knock, a Los Angeles food delivery app, and Half Moon Coffee, a direct-to-consumer roaster.

Pages

Landing and funnel pages built in conversation with an agent, one page per audience: a city, an age group, a set of interests, a cluster of the brand's own customers (from a CSV, HubSpot or Salesforce) or a single lead. The agent studies the brand's live site for its kit (logo, colors, type, buttons), writes the page in HTML and CSS, checks its own render and revises. Pages can be edited by hand, are published at /p/{brand}/{page}, and come with first-party analytics and a heat map.

Demo brands: Tágide Festival, a music festival in Lisbon, and Gwen, a premium credit card.

Email

An AI email builder for campaigns that send a different version to each taste group or person. The agent picks items from the brand's catalog by taste, writes and lays out each version, and checks it on desktop and phone. Sending goes through Resend.

Demo brands: Stackhouse Books, a bookshop, and Covers, a London restaurant booking app.

Concierge

A shopping assistant a store adds with one script tag. It opens with a short taste questionnaire built from Qloo's view of the shopper's city, then recommends from the store's own catalog with a one-line reason for each pick. Catalogs sync from Shopify or WooCommerce, or load from a CSV in the Slice format. The merchant console shows conversations, taste insights across shoppers, and catalog health.

Demo stores: Marlow, home and living, and Fold, multi-brand fashion.

Slice for Mac

A personal agent that lives in the MacBook notch. Hold fn + control and talk. A Gemini Live voice handles the conversation, while a background worker and a computer-use operator do the longer tasks across your apps and screen, with your taste in mind. It connects to Gmail, Google Calendar, Drive, Slack, Notion and more through Composio.

Download the notarized DMG (macOS 15 or later).

How the agents use taste

• Taste is the brief. Before an agent starts, it gets a Qloo read of its audience: the entities and tags that define a city, a demographic, a set of interests, a customer cluster, or a shopper's answers to the questionnaire.
• Qloo is a set of tools. While it works, the agent calls Qloo when it needs more, such as insights for a place, related entities, audience demographics or brand affinities, and decides for itself when and how.
• Every result shows its evidence. Picks, pages, emails and creatives carry the Qloo signals they rest on, so a business can see why the agent chose what it chose.

All Qloo requests go through a cache in Convex (packages/backend/convex/qloo.ts). Responses are stored privately on the server and never committed to this repository.

Architecture

flowchart LR
  shopper(["Shoppers on store sites<br/>(slice.js)"])
  business(["Businesses in the console"])
  mac(["Slice for Mac"])

  web["<b>apps/web</b><br/>Next.js on Vercel"]
  agent["<b>apps/agent</b><br/>Python agents on Cloud Run"]
  convex["<b>packages/backend</b><br/>Convex"]
  vertex["<b>Vertex AI</b><br/>Gemini, Nano Banana, Veo"]
  qloo["<b>Qloo</b><br/>Insights API"]

  shopper --> web
  business --> web
  mac -- "voice over WebSocket" --> agent
  web -- "server-side" --> agent
  web --> convex
  agent --> convex
  agent --> vertex
  convex -- "cached" --> qloo

Loading

• apps/web serves the console, the Concierge UI that slice.js opens on a store, and published pages. It calls the agents server-side only.
• apps/agent runs the Concierge, Pages, Email and Media agents (Google ADK, FastAPI) and the Slice for Mac brain.
• packages/backend holds the data, files and auth checks, and is the only path to Qloo.

Tech stack

Layer	Built with
Web app and demo sites	Next.js 16, React 19, Tailwind CSS 4, shadcn/ui, Clerk
Store widget	TypeScript, Vite
Backend and data	Convex
Agents	Python, Google ADK, FastAPI, Gemini 3.8 Flash on Vertex AI, Playwright
Image and video	Nano Banana 2 and Pro, Veo 3.1
Taste	Qloo Insights API
Integrations	Composio (Shopify, HubSpot, Salesforce, Google Workspace, Slack, Notion), WooCommerce, Resend, Firecrawl, Pexels
Mac app	Swift, SwiftUI, XcodeGen, Gemini Live
Hosting	Vercel, Google Cloud Run
Monorepo	pnpm workspaces, Turborepo

Repository layout

apps/
  web/            Next.js app: landing page, console (/dashboard), concierge UI (/embed), published pages (/p)
  widget/         slice.js, the script tag a store adds
  agent/          Python agents and API server: Concierge, Pages, Email, Media, catalog sync, the Mac brain
  mac/            Slice for Mac, the Swift notch app
  stores/         Demo brand sites: Marlow, Fold, Tágide, Gwen, Knock, Half Moon, Stackhouse, Covers
packages/
  backend/        Convex schema, functions, HTTP routes and the Qloo cache
  qloo/           Typed client for the Qloo API
  demo-catalogs/  Marlow and Fold product catalogs, with validation and export scripts
  demo-data/      Customer datasets and a seed script for the Pages demos
docs/             Design notes, contracts and runbooks

Local development

Prerequisites

• Node.js 20.9 or later and pnpm 10 (corepack enable)
• uv and Python 3.11 or later
• A Convex project, a Clerk application, a Google Cloud project with Vertex AI enabled, and a Qloo API key
• Optional, per feature: Resend (sending email), Composio (Shopify, CRMs, Mac integrations), Firecrawl and Pexels (Pages intake and imagery)
• Optional, for the Mac app: Xcode and XcodeGen (brew install xcodegen)

1. Install

pnpm install

2. Clerk

1. In the Clerk dashboard, open your application, go to API keys, and copy the publishable and secret keys.
2. On the Convex integration page, select the application and click Activate Convex integration. Copy the Frontend API URL it shows (https://<name>.clerk.accounts.dev).

3. Convex

pnpm --filter @slice/backend run convex:setup

Log in, choose existing project and pick your project. This writes packages/backend/.env.local with CONVEX_DEPLOYMENT and CONVEX_URL.

The first push fails until Convex knows the Clerk issuer. Set it from a second terminal and the running setup retries on its own:

cd packages/backend
npx convex env set CLERK_JWT_ISSUER_DOMAIN https://<name>.clerk.accounts.dev

Then give Convex the Qloo key and a shared secret for the agent:

npx convex env set QLOO_API_KEY <your-qloo-key>
npx convex env set QLOO_PROXY_SECRET $(openssl rand -hex 32)

The agent needs the same secret in step 4; read it back with npx convex env get QLOO_PROXY_SECRET.

To send email, also set RESEND_API_KEY and RESEND_WEBHOOK_SECRET.

4. Environment files

cp apps/web/.env.example apps/web/.env.local
cp apps/widget/.env.example apps/widget/.env.local
cp apps/stores/.env.example apps/stores/.env.local
cp apps/agent/.env.example apps/agent/.env

File	Fill in
apps/web/.env.local	NEXT_PUBLIC_CONVEX_URL (the CONVEX_URL from step 3) and the two Clerk keys
apps/agent/.env	GOOGLE_CLOUD_PROJECT, GOOGLE_APPLICATION_CREDENTIALS (a service account key with the Vertex AI User role), CONVEX_URL, CONVEX_SITE_URL and QLOO_PROXY_SECRET
apps/widget/.env.local, apps/stores/.env.local	Nothing; the defaults point at the local servers

The optional keys (Composio, Firecrawl, Pexels, Shopify, WooCommerce) are documented inline in apps/agent/.env.example. The Google credentials and the Qloo key live only in the agent and in Convex, never in the web app or the widget.

5. Run

pnpm dev

This starts Convex, the web app, the widget, the demo sites and the agent server together.

Service	Address
Web app and console	http://localhost:3000
Demo brand sites	http://localhost:3002/marlow (also /fold, /tagide, /gwen, /knock, /halfmoon, /stackhouse, /covers)
Widget playground	http://localhost:5173
Agent server	http://localhost:8000

To load the Marlow and Fold catalogs into Convex:

cd packages/backend && npx convex run seedDemo:seed

Scripts

Command	What it does
pnpm dev	Run every app in development
pnpm build	Production builds
pnpm typecheck, pnpm lint, pnpm format	Type-check, lint and format the whole repo
pnpm --filter @slice/mac app	Build Slice for Mac (Debug) and relaunch it
pnpm --filter @slice/mac release	Build, sign and notarize the Mac DMG
pnpm --filter @slice/demo-catalogs validate	Check the demo catalogs against the schema
cd packages/backend && npx convex run qloo:warm	Pre-cache the Qloo requests the demos use

Adding the Concierge to a store

Create a widget in the console, then paste its snippet into the store's theme:

<script
  src="https://slice-agents.vercel.app/slice.js"
  data-slice-key="<store-key>"
  async
></script>

The console's Install tab also has the snippet for Shopify, Next.js and Tag Manager, limits the widget to the store's own domains, and shows every site it has loaded on.

Deployment

The web app and demo sites deploy to Vercel from main, the agents run on Google Cloud Run, and Slice for Mac ships as a Developer ID signed, notarized DMG. See docs/deploy.md and docs/mac-release.md.

Documentation

Topic	Docs
Deployment	deploy.md, mac-release.md
Pages	pages-plan.md, pages-contracts.md, pages-html-builder.md, pages-crm.md, tagide-eval.md
Email	email-plan.md
Concierge	concierge-stores.md, demo-stores.md, test-taste-profiles.md
Slice for Mac	mac-plan.md

License

MIT