TastePolicy MCP Server

TastePolicy

AI that helps without replacing you.

TastePolicy is personal agency infrastructure for culturally intelligent AI.

Qloo tells an agent what culturally fits a person. TastePolicy answers the next question: may the agent actually go this far with that choice, for this person, in this context, right now?

Margaret's mobility changed. Her taste didn't.

A recommender asks, “Will this person like it?”
Qloo gives grounded cultural evidence for that question.
TastePolicy adds real-world requirements, explicit user rules and an authority ceiling so a strong taste match does not silently become permission.

Start here

• Judge walkthrough: /judge — three short acts, no login.
• Your own experience: sign up, then /onboarding saves your taste, what matters and how far AI may go; /experience and /dashboard use only your saved data.
• Assisted Choice example: /demo/margaret — the labelled live Margaret scenario with real Qloo data.
• Product: /
• Developer platform: /developers
• Trust boundary: /trust
• Live readiness: /status

Production: https://tastepolicy.vercel.app

The core thesis

Taste intelligence is not permission.

The same candidate passes through three distinct layers:

1. Generic AI can produce a plausible suggestion, but it does not have grounded cultural evidence.
2. Qloo resolves cultural identities, discovers cross-domain candidates and provides candidate-specific affinity/explainability.
3. TastePolicy keeps that cultural evidence separate from factual context, the person's requirements, their rules and the action the agent is requesting.

A candidate can be a strong cultural match and still produce ASK or REJECT. Affinity never overrides an explicit boundary.

Four explicit outcomes

Missing evidence is never silently converted into permission. Provider failure is never reported as dislike.

How Qloo is used

TastePolicy does not recreate Qloo's taste graph and does not substitute generic LLM guesses for Qloo evidence.

The current Qloo integration uses:

• GET /search to resolve cultural anchors;
• GET /entities to verify the identities being used;
• GET /v2/insights for cross-domain discovery;
• GET /v2/insights again for candidate-specific affinity/explainability.

The division of responsibility is deliberate:

Without Qloo, TastePolicy intentionally does not pretend a generic model guess is equivalent cultural evidence.

Assisted Choice

The flagship judge story follows Margaret, 72, who still loves intimate live jazz, classic cinema and Italian dining but now needs a step-free entrance, a short walk and a quieter setting.

The public flow:

1. Qloo resolves her five cultural anchors to real entities.
2. Qloo discovers real places from those cross-domain signals.
3. TastePolicy presents the cultural match separately from real-world facts.
4. If a critical requirement such as the entrance is unknown, the answer is ASK, not a guess.
5. After the missing fact is answered, the same deterministic evaluator runs again.
6. A qualifying choice can become ALLOW, but only up to Margaret's saved limit: SHORTLIST.
7. If the agent tries to present that option as the final choice, the authority ceiling forces ASK again.
8. Why this happened keeps Qloo evidence, place facts, requirements, rules and the final decision separately attributable.

A known hard violation becomes REJECT. If the required evidence cannot be obtained, the result is UNAVAILABLE.

Two concise variants demonstrate the same contract beyond Margaret:

• /experience/sensory-aware
• /experience/low-energy

Needs are represented as functional rules such as walking_minutes <= 10 or noise-level constraints, never as diagnoses inferred from cultural taste.

Real place experience

TastePolicy's current open place layer uses:

• MapLibre + OpenFreeMap for the branded journey map;
• OpenStreetMap for explicit place facts where available;
• KartaView and Panoramax for real street imagery, sequences and 360° coverage where available;
• optional openrouteservice/HeiGIT routing;
• licensed Wikimedia or authored Taste imagery as fallback.

Media is presentation, not authority evidence. A nearby street image is labelled as nearby rather than falsely presented as the exact venue. Missing accessibility or route facts remain unknown.

A Google Places adapter also exists as an optional factual provider path, but the public open-place experience does not depend on Google being configured.

One deterministic authority contract

A requirement is represented separately from a candidate fact.

Requirement = {
  key,
  operator,
  expected,
  importance,
  sourceRequirement,
};

CandidateFact = {
  key,
  value,
  state: "known" | "unknown" | "unavailable",
  evidence,
  source,
  sourceType,
  observedAt,
};

Qloo is never accepted as the source of a place-accessibility fact. Cultural evidence, factual evidence and authority remain separate.

Subjective-action ladder

TastePolicy distinguishes a good recommendation from how far the agent may act on it:

The person sets the furthest rung the AI may reach without asking. A request above that ceiling converts an otherwise valid recommendation into ASK.

Deeper authority infrastructure

The simple human experience sits on one shared authority model with:

• versioned Passport/state;
• versioned rules and action ceilings;
• Cultural Receipts;
• bounded leases tied to candidate/context commitments;
• drift checks and revocation;
• signed, selectively disclosed proofs;
• timeline/history;
• REST;
• TypeScript SDK;
• remote MCP;
• A2A;
• OAuth 2.1 / PKCE.

These surfaces extend the same evaluator rather than introducing separate policy engines.

Safety boundary

ALLOW is preference-choice authority only.

It never grants:

• payment or purchase authority;
• booking authority;
• messaging authority;
• account changes;
• private-data disclosure;
• legal authority.

Memory is not permission. Affinity is not authorization. A recommendation is not authorization. Verifying a proof is not authorization.

Local development

Requirements: Node.js 24 and npm.

npm ci
cp .env.example .env.local
npm run dev

Core environment variables:

• QLOO_API_KEY
• QLOO_API_BASE_URL
• QLOO_AFFINITY_SCALE
• DATABASE_URL
• BETTER_AUTH_SECRET
• BETTER_AUTH_URL
• TASTEPOLICY_PUBLIC_ORIGIN
• TASTEPOLICY_ATTESTATION_KEY

Optional place providers/routing are documented in docs/architecture/open-place-providers.md and .env.example.

Apply database migrations with:

npm run db:migrate:all

Verification

npm run verify
npm run test:e2e
npm run conformance:compatible
npm run tastebench:validate

Real-PostgreSQL suites can also run with TASTEPOLICY_TEST_DATABASE_URL configured.

Evidence is labelled by kind — LIVE, HOSTED, CONTROLLED, SYNTHETIC, LOCAL, CI, NOT EXECUTED or BLOCKED — so tests and planned integrations are not presented as production facts.

See:

• docs/evaluation/
• docs/evaluation/m24-execution-ledger.md
• docs/evaluation/m25-execution-ledger.md
• docs/architecture/qloo-contract-review.md
• docs/architecture/open-place-providers.md

Current limitations

• TastePolicy can only detect external drift that an agent or provider presents back to it.
• Open place/accessibility coverage varies by location; missing facts remain unknown.
• Vendor-assistant commissioning status is tracked separately and is not implied by protocol support.
• TasteBench results are engineering/evaluation evidence, not population-level preference or safety claims.

License

Apache-2.0. See LICENSE.