A co-founder who has met your customers.
Every new café, bar or store is designed on gut feel: the founder’s own playlist, a Pinterest board, a menu copied from the place down the road. An LLM makes the same guess faster. TasteLaunch starts from the opposite end — the real, aggregate tastes of the people who already spend money at similar places in that exact city — and designs the business around them.
Why it needs Qloo
A language model knows which artists are famous. It does not know that the people who frequent a city’s specialty coffee shops over-index on a particular indie artist, a particular apparel brand and a particular film, or which part of town that taste concentrates in. Qloo’s taste graph links 250M+ entities across music, film, TV, brands, places and food to real affinity data — that cross-domain jump is the whole product.
One run, step by step
| Step | Call | Detail |
|---|---|---|
| Understand the concept | LLM | Turns “specialty coffee café” into directory categories (coffee shop, cafe). |
| Ground the category | GET /v2/tags | filter.query=<category> · feature.semantic_search=true → real Qloo place tags |
| Map the competition | GET /v2/insights | filter.type=urn:entity:place · filter.location.query=<city> · filter.tags=<tags> · operator.filter.tags=union |
| Read their customers (cross-domain) | GET /v2/insights ×6 | signal.interests.entities=<top 5 competitors> with filter.type = artist, brand, movie, tv_show, podcast, and urn:tag + filter.tag.types=urn:tag:specialty_dish:place |
| City-wide sound | GET /v2/insights | filter.type=urn:entity:artist · signal.location.query=<city> |
| Who they are | GET /v2/insights | filter.type=urn:demographics · signal.interests.entities=<competitors> |
| Where the taste lives | GET /v2/insights | filter.type=urn:heatmap · filter.location.query=<city> · signal.interests.entities=<competitors> → opportunity zones = taste affinity minus competitor density |
| Agent explores (tool calls) | GET /v2/insights · /search | The LLM decides up to 3 extra look-ups, e.g. brands loved by fans of the #1 artist (filter.exclude.entities keeps it fresh) |
| Design + validate | LLM → validator | The kit cites evidence aliases; any pick that doesn’t resolve to a real Qloo entity is removed, and names are rewritten from Qloo’s own records |
| Benchmark | GET /search + /v2/insights | Same model plans without Qloo; its picks are resolved by name and both plans are scored together with filter.results.entities |
A full run is about 15–25 Qloo calls. Every call streams into the trace with its exact parameters; the key never leaves the server.
No hallucinated picks
The model never sees raw IDs. It sees a numbered evidence ledger (A1 = an artist, B3 = a brand, D2 = a dish…) and must cite those aliases. The server validates every citation, drops anything that doesn’t match the right kind of entity, and replaces names with Qloo’s own. If the model is unavailable, a deterministic planner builds the kit straight from the evidence.
Founder–market fit
Add your own favourites (artists, films, brands). The agent resolves them in Qloo and scores each one against the competitors’ audience in a single /v2/insights call with filter.results.entities. High-fit favourites become “founder’s signature” picks; low-fit ones are flagged honestly: you are not always your customer.
Steer the agent
After the first kit, tell it “make it more premium” or “aim it at college students”. The agent decides which extra Qloo look-ups the change needs (a pricier place category, a younger audience’s artists…), then redesigns. Every version stays one click away.
Use it from your own agent (MCP)
TasteLaunch is also an MCP server (Streamable HTTP). Add https://<this-site>/api/mcp as a remote MCP server in Claude, Cursor or any MCP client and your agent gets two tools:
read_local_taste | category + city → competitors, what their customers love across domains, demographic skew and opportunity zones (fast, no LLM needed). |
launch_brand_kit | concept + city (+ founder taste) → the full brand kit in Markdown, with founder–market fit and the with/without-Qloo benchmark. |
The honest caveats
- Affinity describes aggregate taste, not individuals and not sales.
- The benchmark uses Qloo’s own ruler; it shows how far a model working from memory strays from where the taste graph puts this audience.
- Opportunity zones combine taste density with how many matched competitors are within 1.2 km; rent, footfall and licensing still need a human.
- Only public names (business categories, cities, artists, brands) are sent to Qloo. No personal data.