动态血糖数据MCP服务器

Dexcom MCP Server

MCP (Model Context Protocol) server that connects an MCP host (such as Claude Desktop) to personal Dexcom CGM data for human-in-the-loop glucose analysis, event logging, and modeling. The host can analyze and recommend; you decide and act. This project does not automate insulin delivery or change pump settings.

Features

• Glucose tools — latest reading, ranges, daily summaries, and statistics
• Trend analysis — patterns over days/weeks, expected vs actual, parameter-drift signals
• Event logging — insulin, carbs, and exercise, plus timeline / per-type retrieval
• Charts — timeline, daily, weekly, and AGP-style visualizations
• Modeling — baseline ISF/ICR/basal parameters, impact predictions, adaptive insights
• Transports — stdio (local MCP hosts) or http (remote connector with OAuth 2.1 + PKCE)

Requirements

• Node.js 18+ (Docker image uses Node 20)
• Dexcom Developer API credentials (developer.dexcom.com)
• A Dexcom CGM account with data available via the Developer API
• A Turso database (OAuth tokens and persisted readings/events)

Optional: Dexcom Share username/password as a best-effort fallback when the Developer API path fails.

Quick start (local / stdio)

1. Clone and install

git clone https://github.com/cykj40/dexcom-mcp-server.git
cd dexcom-mcp-server
npm install

2. Configure environment

cp .env.example .env

Set at least:

Optional (stdio / bootstrap):

When TRANSPORT=http, also set:

Do not commit .env or any database files.

3. Dexcom OAuth (one-time)

npm run oauth

Complete the browser flow so tokens land in Turso (or use the optional bootstrap env vars once). Manual flow: npm run oauth:manual.

4. Build and run

npm run build
npm start

Development:

npm run dev

5. Claude Desktop (stdio)

Example claude_desktop_config.json (use absolute paths and your own secrets):

{
  "mcpServers": {
    "dexcom": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/dexcom-mcp-server/dist/index.js"],
      "env": {
        "TRANSPORT": "stdio",
        "DEXCOM_CLIENT_ID": "your_client_id",
        "DEXCOM_CLIENT_SECRET": "your_client_secret",
        "DEXCOM_REDIRECT_URI": "http://localhost:3000/callback",
        "TURSO_DATABASE_URL": "libsql://your-db.turso.io",
        "TURSO_AUTH_TOKEN": "your_turso_auth_token",
        "DEXCOM_API_ENV": "production"
      }
    }
  }
}

Restart Claude Desktop after saving.

HTTP transport (optional / remote)

With TRANSPORT=http, the server exposes:

• /.well-known/oauth-authorization-server and /.well-known/oauth-protected-resource
• /authorize and /token (OAuth 2.1 authorization code + PKCE S256, owner-key approval)
• Authenticated /mcp
• Public /health

Operator CLI (after build): node dist/auth/admin.js list and node dist/auth/admin.js revoke <grant-id>.

Prefer stdio for local use. Treat HTTP mode as a high-sensitivity deployment: strong secrets, exact redirect allowlists, HTTPS issuer URL, and private hosting.

Optional deploy config lives in fly.toml (Fly.io). Do not publish live app URLs or secrets in docs.

Available MCP tools

Glucose

• get_latest_glucose
• get_glucose_range
• get_daily_summary
• get_glucose_statistics

Analysis

• analyze_trends
• compare_expected_vs_actual
• detect_parameter_drift

Events

• log_insulin, log_carbs, log_exercise
• get_event_timeline
• get_insulin_events, get_carb_events, get_exercise_events

Charts

• generate_chart

Modeling

• get_baseline_parameters
• update_baseline_parameters (requires explicit confirmation in the tool args; treat as sensitive)
• predict_glucose_impact
• get_adaptive_insights

Example prompts (generic)

What's my latest glucose and trend?

Summarize yesterday's glucose statistics.

Log 5 units of rapid insulin for lunch and 45g carbs.

Show an event timeline for the last 24 hours.

Predict the glucose impact of 40g of carbs with my current baseline.

Data storage

Runtime persistence uses Turso (TURSO_DATABASE_URL / TURSO_AUTH_TOKEN): OAuth tokens, glucose readings, events, and related records. This is a third-party hosted database you control via your Turso account.

Do not commit database files; keep data/ and *.db* gitignored. Prefer keeping the repository private while it holds health-data tooling.

Security notes

• Keep credentials in environment variables or a secret store — never in source.
• Prefer TRANSPORT=stdio for local use.
• HTTP mode uses OAuth 2.1 + PKCE with owner approval and hashed tokens. Rotate any legacy MCP_AUTH_TOKEN if it may have been exposed, and prefer OAuth grants.
• This server is assistive and must not be treated as a closed-loop controller.
• Review git history before making the repo public: do not leave historical database blobs or secrets reachable.

npm scripts

Project structure

dexcom-mcp-server/
├── src/
│   ├── auth/         # HTTP OAuth, crypto helpers, grant admin CLI
│   ├── config/       # Env validation
│   ├── db/           # Turso client, migrations, queries, token store
│   ├── services/     # Dexcom API/Share, glucose, events, modeling, charts
│   ├── tools/        # MCP tool registration
│   ├── types/        # Shared types
│   ├── utils/
│   ├── oauth-helper.ts
│   ├── manual-oauth.ts
│   └── index.ts      # Entrypoint (stdio / http)
├── test/             # Vitest tests and fixtures
├── .env.example
├── fly.toml          # Optional Fly.io deploy config
└── package.json

Known limitations / missing information

• Share API path is undocumented / best-effort and may be unreliable
• Baseline updates currently rely on a caller-supplied confirmation flag — treat compromised clients carefully
• Schema may seed placeholder physiology defaults on empty DB; bootstrap intentionally for real use
• No CI workflows documented in this README
• Health data is sensitive — protect the repo, secrets, remote endpoint, and any clones

Medical disclaimer

Assistive tool only — not medical advice and not FDA-approved medical software.
You are responsible for all treatment decisions. Consult your clinician before changing insulin or other therapy. Never rely solely on this tool for dosing.

License

MIT — see LICENSE.