loupe

Appraise the real value. One canonical, sourced page per GLP-1 / cardiometabolic drug — real acquisition cost, live FDA shortage status, verbatim label facts, and same-class alternatives — unified from five free, keyless, public-domain US-government datasets. Built agent-native: deep-links, copy-as-API on every panel, and an MCP server return the same Drug contract the human UI renders.

Not medical advice. loupe is not affiliated with the FDA, NLM, or CMS. Drug-PRODUCT reference data only — zero PHI.

What it does

loupe unifies five public-domain US-government drug datasets into one sourced page per GLP-1 / cardiometabolic drug: the NADAC acquisition cost, current FDA shortage status, verbatim label facts, and same-class alternatives — each value carrying its source and as-of date.

Related MCP server: drug-pipeline-mcp

Guardrails

A guardrail that isn't enforced by the type system, a runtime guard, and a test does not exist.

1. Never dose-advise, never recommend a switch — alternatives are clinician-gated.

2. Cost ≠ copay — NADAC is acquisition cost; every price carries cost_basis.

3. FAERS = reports, not rates — never an incidence rate, never a denominator.

4. Labels are quoted verbatim with source + as-of date. No paraphrase, ever.

5. Honest loading / empty / error / stale states on all data.

6. Explicit "not medical advice / not affiliated" disclosure across UI, API, and MCP.

Architecture — one contract, three faces

A build-time prefetch (scripts/prefetch.ts) fans out to all upstreams, validates each record with Drug.parse(), and commits a static catalog (src/data/snapshot/*.json). The SPA, the Hono read API, and the MCP server all read that one Drug contract (src/core/schema.ts) — nobody live-hits the upstreams at request time, so the demo is deterministic and no rate-limit budget bites. Snapshots omit the guardrail literals; Drug.parse() back-fills them on read, so a snapshot can never ship a tampered disclaimer (enforced by src/core/literals.test.ts).

upstreams ──prefetch──▶ src/data/snapshot/*.json ──▶ Drug (src/core/schema.ts)
  openFDA ×4                                            │      │       │
  RxNorm/RxClass                                      SPA   Hono /v1  MCP (stdio)
  CMS NADAC + Part D                            (bundled)  (agents)  (agents)
  DailyMed                                  toApiCall(DrugQuery) → curl/python/ts + deep-link

Run it

npm install
npm run prefetch     # rebuild the catalog from live upstreams (committed; optional)
npm run dev          # SPA  → http://localhost:3001
npm run api          # API  → http://localhost:8787  (also serves the built SPA)
npm run mcp:smoke    # spawns the MCP server and runs the "can I afford Zepbound?" example
npm test             # vitest: contract, snapshot-literal invariant, guardrail-render, UI
npm run test:coverage # the above + a coverage gate on the request-time logic (CI-enforced)
npm run build        # tsc --noEmit && vite build

Agent surfaces

• Read API (keyless): GET /v1/drugs, /v1/drugs/{slug}, /v1/drugs/{slug}/{panel}, /v1/search?q= · OpenAPI at /openapi.json · docs at /docs.

• Discovery: /llms.txt and /.well-known/navigator.json.

• MCP (stdio) — add to an MCP client:

  { "mcpServers": { "loupe": { "command": "npx", "args": ["tsx", "src/mcp/server.ts"] } } }

  Tools: get_drug, list_drugs, search_drugs, get_drug_cost, get_drug_shortage, get_drug_label, get_drug_alternatives.

Deploy

One container serves both faces (the SPA bundles its own catalog; the API serves /v1):

docker build -t loupe . && docker run -p 8787:8787 loupe
# → http://localhost:8787  (SPA at /, API at /v1, docs at /docs)

Snapshots are committed, so the image builds with no network. Refresh them on a schedule by running npm run prefetch in CI and redeploying.

Example

/?drug=tirzepatide-zepbound&view=cost opens the cost view for Zepbound. The same data is available without the UI — GET /v1/drugs/tirzepatide-zepbound/cost, the get_drug_cost MCP tool, or a panel's "copy as API" snippet — each returns the same Drug projection.

Key decisions

• Build-time static catalog, not live per-request fetch — fixes deterministic demo + the openFDA 1000/day shared-IP wall + Lambda cold-cache in one move.

• est. $/mo only when it needs no clinical assumption (oral, dose-countable). For mL-priced injectables it's withheld; the factual package cost is shown instead.

• Approval = "an SPL exists" + has_boxed_warning — we never algorithmically infer off-label/withdrawn (that would manufacture a regulatory claim).

• NDC normalized to 11-digit 5-4-2 before the NADAC join (classic silent-fail point).

See docs/SPEC.md for the full data contract and rationale.

License

Apache License 2.0. The underlying datasets are public-domain US-government works (FDA, NLM, CMS); loupe is not affiliated with those agencies.