Taste Overlap
Two people, a few favourites each (artists, films, shows, books, places). Qloo's taste graph ranks places in your city for person A, for person B and for both together. Taste Overlap keeps only the places that rank well for both (harmonic mean of the two affinities, so one-sided picks drop out), explains each pick with the shared taste tags, and orders them into an evening: something to do, food, then a drink.
Without Qloo there is nothing to show: the overlap is computed from Qloo's per-person affinities, not from an LLM guess.
Run
Needs Node.js 22.19+ and the official Qloo harness (the supported event surface):
Demo without credentials: QLOO_TRANSPORT=demo npm start replays a real recorded run (demo/vienna.json, 11 qloo exec calls) through the same code. Only that example works (Vienna; Radiohead + Wes Anderson / Taylor Swift + Studio Ghibli); other inputs return an error instead of invented places.
Optional: QLOO_TRANSPORT=rest with QLOO_API_KEY uses the direct REST client (src/qloo.js) instead; QLOO_BASE, PORT.
Use it from an agent (MCP tool)
Remote, nothing to install: the live demo also speaks MCP at https://taste-overlap.onrender.com/mcp (streamable HTTP, JSON answers). In claude.ai or the Claude app: Settings → Connectors → Add custom connector and paste that URL. Tool calls share the demo's per-visitor and daily limits.
Local (stdio):
Taste Overlap is also an agent tool: mcp-server.js is a dependency-free MCP stdio server with one tool, plan_shared_evening (city, person_a[], person_b[]). It returns the explained evening as text plus structured content, and any MCP client can call it:
Without a key, QLOO_TRANSPORT=demo makes the tool answer the recorded Vienna example.
Test
How it works
src/mcp.js (default transport) - one long-running qloo mcp process (the harness's canonical MCP adapter) serves every workflow as qloo_<op>; on a small host this takes ~0.5-1.2 s per call instead of >30 s for a fresh qloo exec. QLOO_TRANSPORT=exec switches back to one qloo exec per call.
src/harness.js - the Qloo workflows: recommend (target_type: place, the favourites as signals, filter_location: the city) for A, B and both; entity_tags for the tags that explain each pick. The harness resolves the names and validates input before calling Qloo; when a name is ambiguous (status: needs_input) the client takes Qloo's top-ranked non-album candidate, retries once with its entity id and returns that choice. Workflow names are fixed, user text goes in on stdin.
- Each pick's reason comes from Qloo's per-result
explainability (signal.interests.entities): which favourites drove the place and by how much.
src/demo.js - records and replays real harness envelopes for the credential-free demo.
src/qloo.js (fallback) - resolves each favourite with /search, ranks places with /v2/insights (filter.type=urn:entity:place, signal.interests.entities, filter.location.query). The key is sent as a header from the server and never reaches the browser.
src/overlap.js - pure logic: shared picks, reasons, evening order (unit-tested). Qloo returns at most 20 places per call, and two tastes can share none of them (measured: Berlin, Daft Punk + Christopher Nolan vs Adele + Pixar). Real overlaps come first; the rest is topped up from the joint list and labelled "combined taste". A place scored one-sided in both personal lists never comes back that way.
src/plan.js - three Qloo calls (A, B, both) -> overlap -> explained plan.
server.js - dependency-free Node server, serves public/ and POST /api/plan.
mcp-server.js - the same plan as an MCP tool for agents; src/transport.js picks the Qloo transport for both.
License
MIT