featherless

featherless

TypeScript SHARP-on-MCP-compliant FHIR R4 MCP server with MCP-UI clinical dashboards. Runs on Cloudflare Workers.

Sibling project to sharp-fhir-mcp (Python, FastMCP 2.x). Same SHARP semantics, same tool set, same MCP-UI Chart.js visualizations — re-implemented in TypeScript for the edge.

What's included

Core

• SHARP HTTP transport via Cloudflare agents SDK (McpAgent, Durable Object–backed sessions)
• Header-based context (X-FHIR-Server-URL, X-FHIR-Access-Token, X-Patient-ID) with SMART JWT claim fallback
• Strict and permissive context modes
• experimental.fhir_context_required capability injected into initialize

Tools (parity with sharp-fhir-mcp)

• tools/fhir.ts — generic FHIR R4 search/read
• tools/clinical.ts — Patient / Encounter / Appointment / Allergy / Medication / Problem
• tools/lab-imaging.ts — Observation / DiagnosticReport / DocumentReference
• tools/clinical-context.ts — aggregated visit context with derived alerts
• tools/memory.ts — clinical memory backed by Cloudflare Vectorize + Workers AI + D1
• tools/visualization.ts — MCP-UI Chart.js dashboards (visualize_lab_trend, visualize_vitals, visualize_patient_dashboard)

Getting started

pnpm install
# create the Vectorize index (once)
pnpm wrangler vectorize create featherless-memory --dimensions=768 --metric=cosine
# create the D1 database (once), put the id into wrangler.jsonc
pnpm wrangler d1 create featherless-memory-meta
pnpm db:migrate:remote
# develop
pnpm dev

SHARP context

Forward these headers on every tools/call:

License

MIT.