Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@nightscout-mcp how was my time in range this week?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Ask Claude about your Nightscout CGM data. Read-only — it cannot change anything in Nightscout.
Not a medical device. Don't make treatment decisions from it. Readings can be stale, missing or wrong, and an LLM can misread them.
(Illustrative session — invented numbers, not a capture of anyone's data.)
Point your client at your deployed server and send a bearer token. Standing the server up is further down.
Claude (web, desktop, mobile) — Customize → Connectors → Add custom connector
field | value |
URL |
|
header |
|
value |
|
Include the word Bearer and the space — Claude sends the value verbatim and
adds no prefix of its own. Request-header auth is a Claude beta; if you don't
see a Request headers section, ask Anthropic for access.
Claude Code
claude mcp add --transport http nightscout https://your-host.example.com/mcp \
--header "Authorization: Bearer <your MCP_BEARER>"Anything else that speaks MCP — this is a standard streamable-HTTP server.
Give it the /mcp URL and an Authorization: Bearer … header. Verified with
Claude, Claude Code and FastMCP's own client; nothing in it is Claude-specific.
Related MCP server: MCP Nightscout
Right now
Tool | Returns |
| Latest reading, trend direction, and how old it is |
| Readings over the last N hours (default 3) |
| Active insulin and carbs on board (Loop/AAPS) |
| Pump reservoir and battery, uploader battery, loop health |
| Age of cannula, sensor, insulin and pump battery |
Over time
Tool | Returns |
| Low / in-range / high split, average, GMI, CV% (default 24h) |
| Glucose by hour of day — when you actually run low or high |
| The last N days against the N before, with deltas |
| Boluses, carbs and site changes (default 12h) |
Interactive
Tool | Returns |
| The overview as a chart where the client supports it, JSON everywhere |
Built on MCP Apps, so the tool returns data and the host renders the view in a sandboxed iframe. That split matters: the same JSON the chart draws from is what the model reads, so Claude can answer questions about the chart. An image it cannot see would make the conversation worse, not better.
The range buttons don't fetch anything — they ask the host to call the tool again. The iframe holds no credential and never talks to Nightscout. There is also no charting CDN: the view is ~120 lines of hand-written HTML with an empty CSP, because a third party in the render path for glucose data is not a dependency worth taking for rounded corners.
Requires a client that supports the MCP Apps extension. Everywhere else the same tool returns the same JSON and nothing is lost.
Reference
Tool | Returns |
| Basal rates, ISF, carb ratio, targets |
| Nightscout version and configured thresholds |
glucose_patterns is the one worth knowing about. "What is my glucose" is a
question a graph already answers; "what time of day do I reliably go low" is
one it doesn't, and it's the kind of thing a model is good at picking out of
two weeks of readings. Hours are binned in your Nightscout profile's own
timezone, so 3am means your 3am.
There are no write tools. The server cannot change anything in Nightscout.
Nightscout's bolus wizard preview (bwp) returns a suggested insulin
amount. Surfacing it would turn this from something that displays your data
into something that suggests treatment — a different thing entirely, both for
you and for how such software is regulated. Same reasoning for anything that
computes a correction. Every tool here is a read.
This serves personal health data over the public internet, so the defaults are deliberately strict:
A bearer token is required, and the server refuses to start without one. There is no "unauthenticated if you forget the variable" mode — that is the failure a deployment never notices.
Least privilege upstream. It reads Nightscout with a readable access
token, never the admin API_SECRET. Even a total compromise of this server
cannot write to your Nightscout.
Constant-time token comparison, so response timing can't be used to recover the bearer a character at a time.
Errors are sanitised. Nightscout authenticates by query parameter, so the token is in the request URL; upstream failures are re-raised naming only the path and status, keeping the credential out of logs and client-visible errors.
Forwarded headers are trusted only from proxies you name, via
TRUSTED_PROXY_IPS. Unset means trust nobody.
Runs as a non-root user in the container.
/health is the one unauthenticated route. It returns a fixed {"ok": true}
and reads nothing.
In Nightscout: Admin Tools → Subjects → Add. Give it the readable role
only. You get a token like claude-1a2b3c4d5e6f7890.
Do not use your API_SECRET. Nothing here writes, so handing it write
access buys you nothing and costs you everything if the token leaks.
cp .env.example .env
# edit .env — at minimum NS_URL, NS_TOKEN, MCP_BEARERGenerate the bearer with something you didn't invent yourself:
python3 -c "import secrets; print(secrets.token_urlsafe(32))"docker build -t nightscout-mcp .
docker run -d --name nightscout-mcp --env-file .env -p 8787:8787 nightscout-mcpOr without Docker:
pip install -r requirements.txt
set -a && source .env && set +a
python server.pyThe server speaks plain HTTP and expects TLS to be terminated in front of it (Traefik, Caddy, nginx, Cloudflare Tunnel — anything). Claude connects from Anthropic's servers, not from your device, so the URL must be reachable from the public internet.
If your proxy sets X-Forwarded-*, name it:
TRUSTED_PROXY_IPS=172.18.0.0/16In Claude: Customize → Connectors → Add custom connector
URL: https://your-host.example.com/mcp/
Open Request headers, add authorization with value
Bearer <your MCP_BEARER>
Enter the value including the word Bearer and the space — Claude sends the
header verbatim and adds no prefix of its own.
Request-header auth is a Claude beta. If you don't see a Request headers section, ask Anthropic for access. Claude Code can use the same server today via its own MCP configuration.
NS_UNITS=mg/dl (default) or mmol. This only affects how values are
presented; Nightscout is always read in mg/dL and converted on the way out.
This is an independent project. It is not affiliated with, endorsed by, or part of the Nightscout Foundation.
Nightscout itself (cgm-remote-monitor) is licensed AGPL-3.0. This server contains no Nightscout code and links no Nightscout library — it only makes HTTP requests to a running instance's REST API. Consuming an API at arm's length does not create a derivative work, and the AGPL's network clause is conditioned on modifying the covered program, which this does not do. So Nightscout's copyleft does not extend here, and this project is separately licensed.
If you fork this and vendor any Nightscout source into it, that reasoning stops applying to your fork.
MIT — see LICENSE.
Chosen for reach: this is a small connector whose value is that anyone can run it in ten minutes. Copyleft would protect nobody in the common case, which is one person self-hosting it for their own data.
This server cannot be deployed
Your Nightscout glucose data, conversational.
Glucose readings from your LibreLink Up sensor: graph, logbook, stats and summaries (read-only). Sec
Your own SLATE health data in your AI assistant: read, trend, plan and log. Never suggests a dose.
Ask questions in plain language, get answers from your business database. No SQL required.