sikcapri/superglookoquery

**Talk to Claude Desktop about your Glooko diabetes data, whatever pump or CGM combination you're actually running.**
> [!IMPORTANT]
> **Not medical advice.** This is for understanding your own data and preparing sharper questions for your care team. It is not a medical device, and nothing it says should change your therapy on its own. Full text in the disclaimer at the bottom.
> [!NOTE]
> **This started as a fork of Richard Hall's PodQuery** (MIT licensed), which was built around one specific setup: an Omnipod 5 syncing to Glooko. Credit to Richard for the original MCPB packaging and clinical-analysis core. What's different here: the ingestion layer no longer assumes a fixed set of fields belonging to one device. It looks at what your own account's data actually contains and builds its tool list from that. See How device support actually works below for the mechanics.
> [!NOTE]
> **v0.2.0.** Still built and tested end to end against one real setup so far: a Ypsomed YpsoPump running CamAPS FX as the AID app, paired with a Libre 3+ CGM. Other combinations grow in as real accounts contribute data, not by guessing at fields nobody's verified.
---
Contents
* What this actually does
* How device support actually works
* Who this is for
* Privacy, in plain terms
* The clinical auditor persona
* Installing it
* Configuring it
* Using it
* Moving from sample data to your own
* Troubleshooting
* Get in touch
* Tool reference
* How the code is laid out
* Building the .mcpb yourself
* Running the tests
* License
* Disclaimer
---
What this actually does
SuperGlookoQuery connects to your Glooko diabetes data and hands Claude a set of analytical tools, not a raw data dump. Install the extension, ask a question in plain language, and Claude reaches into your history to answer it, live, in the conversation.
People ask things like:
* "How was my time in range last month?"
* "Why do I keep going high in the evenings?"
* "Show me my worst day and tell me what happened."
That means, under the hood:
* **Summaries on demand.** Time in range, GMI, variability, best/worst days and hours, basal/bolus split, over whatever window you ask about.
* **Full-resolution CGM data kept locally.** Every 5-minute reading is archived so nothing's missed, though Claude is steered toward aggregates first and only pulls raw readings when a question genuinely needs them.
* **A real chart, not a description of one.** A clinical-report-style glucose chart opens directly in your browser: colour-coded ranges, hoverable bolus markers, a day-by-day breakdown.
* **Bolus context, not just numbers.** Each bolus is matched against the ISF, carb ratio, and target active at the time, including split or extended deliveries where a pump reports them, so Claude can actually judge whether a dose made sense.
* **Basal behaviour as states.** Suspended, running at max, or blind after a lost signal, rather than raw delivery units.
* **Device-specific extras, only when earned.** Something like CamAPS FX's own operating-mode breakdown appears once your account's data has actually produced that field, not before.
You never call any of this directly. Claude does it mid-conversation, as many small calls as the question actually needs.
---
How device support actually works
Glooko doesn't publish a field-level schema, and different pump/CGM pairings send back genuinely different shapes of data. The original PodQuery handled this by hardcoding one device's fields, which made sense given it was built around its own author's Omnipod 5. This fork takes a different approach:
* **Nothing gets thrown away.** Every field Glooko sends is kept, not just the ones a tool currently reads. Anything not yet mapped to a typed column lands in an extra overflow column instead of being discarded.
* **Tools are earned, not assumed.** A device-specific tool, like the CamAPS pump-mode breakdown, only registers once your own account's data has shown the field it needs at least once. There's no config file to hand-edit and no tool silently misbehaving on hardware it was never written for.
* **You can contribute your own device's shape, from inside the chat.** Ask Claude to review your data for the schema registry, or call get_registry_contribution_report yourself. It builds a report of which fields your account populates and how often, with every real value already swapped for a fixed, fabricated placeholder before you ever see it. You review it, then type the confirmation phrase it gives you yourself (this is deliberately not something Claude can do on your behalf), and submit_registry_contribution runs an independent second privacy scan plus a content-integrity check before opening a pull request against this project's schema-registry/ folder. Your actual data never leaves your machine at any point in this. A node src/submit-registry-entry.js CLI also exists for local development, but it needs a terminal a real Claude Desktop install doesn't have, so the in-chat tools above are the real path for an actual user.