TL;DR: In July 2026 Samsung Health pushed a consent screen asking to use activity, medication, health-record, and menstrual data for AI training — with an opt-out warning so aggressive Samsung had to publicly walk it back a week later. If that was your cue to leave, the FOSS stack is real but has one hard catch: Gadgetbridge does not support Wear OS, so your Galaxy Watch itself can’t come with you.

	Gadgetbridge	Home Assistant	Nextcloud (Tables)
Best for	Band/watch sync with zero cloud	Aggregating scales, watches, sensors	People already running Nextcloud
License	AGPL-3.0	Apache 2.0 (core)	AGPL-3.0
Hardware cost	$0 (your Android phone)	~$80–120 (Raspberry Pi class)	Existing server or Pi
The catch	No Wear OS — Galaxy Watch 4+ excluded	Setup effort; integrations vary by brand	The dedicated Health app is unmaintained

Honest take: If you’re buying new, a Gadgetbridge-supported watch (Garmin, Amazfit, Bangle.js, PineTime) plus Home Assistant is the cleanest privacy-first stack. If you own a recent Galaxy Watch, there is no full FOSS escape hatch today — your realistic move is opting out of AI training and limiting what Samsung Health collects.

What Samsung actually did in July 2026

In early July 2026, Samsung Health began showing a screen titled “Consent to the Use of Health Data for AI Training and Modelling.” The scope was broad: activity data, medication, health records, and menstrual-cycle data, with human review of samples included. Worse, the dialog shown when users tried to decline warned that opting out would delete their Samsung Health data from Samsung’s servers and stop cloud sync — which reads as “consent or lose your history.”

The backlash was immediate, and on July 14 Samsung clarified: only the copy of data retained specifically for AI training is deleted when you opt out. Your health records stay, tracking keeps working, and account sync continues. The original warning was, in Samsung’s own framing, misleading.

So the worst version of the story didn’t hold. But the episode is still a useful forcing function, for two reasons. First, the default is opt-in pressure on some of the most sensitive data a phone collects — menstrual and medication data are exactly the categories you don’t want in any training corpus, anonymized or not. Second, you only found out what declining actually did because journalists pushed. A closed app plus a closed cloud means the vendor’s dialog box is your only documentation. That’s the structural problem self-hosting fixes.

The catch nobody leads with: your Galaxy Watch can’t go FOSS

Most “ditch Samsung Health” advice hand-waves this, so let’s put it first. Gadgetbridge, the standard FOSS companion app for wearables, explicitly does not support Wear OS devices. That excludes every Samsung Galaxy Watch from the Watch4 onward, and the older Tizen-based Gear watches aren’t supported either. There’s a long-standing open discussion about general Wear OS support in the project tracker, but as of August 2026 it is not a thing you can install and use.

The consequence: if your health data currently flows from a modern Galaxy Watch, there is no drop-in open-source pipeline for it. Your options are damage control (opt out of AI training, disable what sync you can, keep using the watch) or changing hardware. That’s an honest cost, and whether it’s worth paying depends on how much the July episode bothered you.

Option 1: Gadgetbridge — zero-cloud sync for supported wearables

Gadgetbridge is an Android app (AGPL-3.0, developed on Codeberg at Freeyourgadget/Gadgetbridge, installable from F-Droid) that pairs directly with wearables over Bluetooth and stores everything in a local database on your phone. No account, no cloud, no vendor app. Version 0.93.0 shipped August 12, 2026 — the project is very much alive.

What it supports instead of Wear OS:

• Amazfit and Xiaomi bands/watches — the historical core of the project, broad coverage.
• Garmin watches — support has expanded substantially; many current models sync activities, heart rate, and sleep. Check the device list on gadgetbridge.org for your exact model before buying, because coverage varies per device.
• Bangle.js and PineTime — fully open-source watches (open firmware, open companion). Fewer sensors and less polish than a Garmin Fenix 7S, but the only tier where the entire stack, firmware included, is auditable.

What you give up: no web dashboard, no automatic multi-device sync, and analysis is limited to what the app charts. Gadgetbridge can auto-export its database, which is the hook for the next layer.

Option 2: Home Assistant — the aggregation layer

Home Assistant (Apache 2.0 core) is the strongest FOSS answer to “where does all my health data live?” It runs on a Raspberry Pi 5 or any spare box, and stores history locally in SQLite by default (InfluxDB if you want long retention).

Relevant integrations, verified against the official integration index:

• Withings — official integration. Scales, sleep mats, and watches sync via the Withings API. Note the threat-model asterisk: data still transits Withings’ cloud; HA pulls a copy locally.
• Apple Health — no official integration, but the established pattern is an iPhone Shortcuts automation that POSTs metrics to a HA webhook. Fully local delivery, some assembly required.
• Garmin Connect — available as a community HACS integration if you keep Garmin’s cloud, or skip the cloud entirely by pairing the watch with Gadgetbridge and exporting into HA.
• Bluetooth LE sensors — heart-rate straps and some scales can be read directly, no vendor anything.

Effort level: this is a weekend project, not an app install. The payoff is one local dashboard for weight, sleep, activity, and anything else your house measures — with retention you control.

Option 3: Nextcloud — with a correction to the usual advice

Lots of 2024-era guides recommend the Nextcloud Health app. Don’t, at least not as your primary plan: the app’s own store page notes it hasn’t been tested against the last three major Nextcloud releases, and the project is openly looking for a new maintainer. Recommending unmaintained software for health data would be malpractice.

The maintained path inside the Nextcloud ecosystem is Nextcloud Tables for storing metrics (weight, workouts, medication logs) plus Nextcloud Analytics for charts — the Health app’s maintainers themselves point users there, and CSV export/import covers migration. It’s spreadsheet-grade rather than purpose-built, but it’s AGPL, self-hosted, and alive. If you already run Nextcloud for files and calendars, it’s a reasonable home for manually tracked or exported health data.

Where local AI fits (and where it doesn’t)

The irony of Samsung’s pitch is that “AI insights from your health data” is genuinely useful — the objection is where the training happens. If you’ve got the FOSS pipeline above, you can point a local model at your own exports with nothing leaving your machine: a small instruct model via Ollama summarizing weekly trends from Gadgetbridge CSV exports, or AnythingLLM answering questions over your logs. An n8n + Ollama workflow can automate the weekly summary entirely. None of this needs a big GPU — 7B-class models handle tabular summaries fine, and if you want to experiment with larger models before buying hardware, a rented GPU on RunPod keeps the training-data question moot because you control the instance.

What no FOSS stack replicates yet: Samsung’s coached experiences — adaptive sleep coaching, guided workout plans. The open ecosystem gives you your data and basic charts, not a product team’s polish. Anyone telling you otherwise is selling something.

The realistic verdicts

• You own a Galaxy Watch 4/5/6/7/8: stay, but opt out of AI training (your history survives, per Samsung’s July 14 clarification) and audit sync settings. Revisit when you next replace the watch.
• You’re buying a wearable now and privacy is the point: Garmin (check Gadgetbridge’s device list first) or Bangle.js/PineTime if you want the fully open stack, synced to Gadgetbridge, aggregated in Home Assistant.
• You already run a homelab: add the Withings integration or a webhook pipeline to Home Assistant this weekend. It’s the highest-leverage single move.
• You mostly log manually: Nextcloud Tables. Skip the unmaintained Health app.

For the hardware side of running all this — Pi vs. mini-PC vs. repurposed desktop, and the GPU question if you want local AI on top — see the self-hosted privacy angle at runaihome.com, and our own minimum viable privacy stack for the AI layer. If your concern extends to code leaving your machine, aicoderscope.com covers local-first AI coding tools.

FAQ

Did Samsung really delete data from users who opted out of AI training? No. The original opt-out dialog implied server data deletion and loss of cloud sync, but Samsung clarified on July 14, 2026 that only the separate copy collected for AI training is deleted. Health records and sync continue working after opting out.

Can Gadgetbridge replace the Galaxy Wearable app for my Galaxy Watch? No. Gadgetbridge does not support Wear OS devices, which includes every Galaxy Watch since the Watch4. It supports Amazfit/Xiaomi wearables, a growing list of Garmin watches, and open hardware like Bangle.js and PineTime.

Is Home Assistant health data really local? The storage is — history lives in a local SQLite (or InfluxDB) database on your own hardware. But cloud-polling integrations like Withings still send your data through the vendor’s servers first; only direct Bluetooth sensors and webhook pipelines are end-to-end local.

Sources

• 9to5Google — Samsung Health won’t delete all of your data if you opt out of AI training after all
• Cybernews — Opt out of Samsung AI training, lose health data
• Gadgetbridge — official site and supported-device list
• Nextcloud App Store — Health (maintenance status)
• Home Assistant — Withings integration docs

Recommended Gear

• Garmin Fenix 7S — multisport watch from Gadgetbridge’s best-supported non-open brand; verify your exact model on the device list.
• Raspberry Pi 5 8GB — enough to run Home Assistant plus a Nextcloud instance for the whole stack above.