DentOps Copilot

An MCP-native agentic copilot for a dental practice. Given a dental X-ray + a patient record, a Claude agent orchestrates several MCP tools to detect findings, pull history, draft a patient-friendly summary and an insurance pre-authorisation narrative, and propose a recall — streaming every step live to a React dashboard over WebSocket. The MCP server (FastAPI + FastMCP) is callable by any MCP client, including Claude Desktop.

Portfolio prototype. Not a medical device. Every generated artifact is a draft for dentist review.

What I built vs. what's third-party

• Mine: the FastMCP server (tools / resources / prompts over JSON-RPC), the Claude agent that orchestrates them through an MCP client, the grounded draft generation, the real-time WebSocket UI, and the deploy setup.
• Third-party / mock: the imaging detector. It sits behind a DetectorBackend interface; the default MockDetector returns fixture findings. No model is trained here. Swap in an off-the-shelf detector via the interface.

Quickstart

cp .env.example .env          # add your ANTHROPIC_API_KEY
make install
make seed
make dev                      # API (uvicorn) + frontend (vite)
# open the printed localhost URL, click "Run demo · patient #1042"

Architecture, tool contracts, and the event protocol are in SPEC.md. Build/finish notes for Claude Code are in CLAUDE.md. Deployment is in docs/DEPLOY.md.

Limitations & safety

Mock patient data and fixture detector output. No clinical validity; outputs are drafts only and route low-confidence findings to a dentist. Do not use with real patient data.