Taste Layer for agents

Agents are good at reasoning and bad at taste. This project gives any agent a taste layer built on the Qloo taste graph, as:

1. An MCP server (mcp/server.ts) with two tools: resolve_taste_anchor and recommend_for_group.

2. A hosted demo (Next.js) that answers the same question twice, once from a plain LLM and once grounded in Qloo, side by side.

The key idea is group coverage ranking. Qloo is queried once per person from what that person likes, plus once for the whole group. Picks are ranked by coverage (how many people the pick fits), then affinity. A plain LLM has no per-person signal, only stereotypes.

Status

Built for the Qloo Agentic Hackathon 2026. Without QLOO_API_KEY the app runs in clearly labeled SAMPLE DATA mode (local fixtures, not Qloo output) so it can be tested offline. With a key it calls Qloo's /search and /v2/insights endpoints on the hackathon base URL. If a Qloo call fails at runtime (expired or revoked key, outage), that request is rerun against the same clearly labeled SAMPLE DATA fixtures instead of erroring; later requests retry the live API. The MCP server behaves the same way.

Related MCP server: Tastecraft

Run

cp .env.example .env.local   # add QLOO_API_KEY (hackathon key), optional LLM_API_KEY
npm install
npm run dev                  # demo at http://localhost:3000
npm test                     # unit tests
npm run mcp                  # MCP server over stdio

MCP client config:

{ "mcpServers": { "taste-layer": { "command": "npx", "args": ["tsx", "mcp/server.ts"], "env": { "QLOO_API_KEY": "..." } } } }

Layout

• src/lib/qloo.ts Qloo client (search + insights, X-Api-Key header)

• src/lib/taste.ts anchor resolution and group coverage ranking

• src/lib/plan.ts runs the with/without-Qloo comparison

• src/lib/sample.ts offline fixtures (sample mode only)

• mcp/server.ts MCP server

License

MIT