Drei Dokumente: QUICKSTART.txt → first-time setup in a few minutes README_APP.md → installation, security, technical setup (this document) USER_GUIDE.txt → all features in daily use, troubleshooting
Garmin Connect is still required — the app pulls data from there via API. This tool does not replace Connect, the Garmin app, or your device sync.
Two versions are available:
| Standard | Standalone | |
|---|---|---|
| Python required | Yes | No |
scripts/ folder needed |
Yes | No |
| First launch speed | Normal | Slightly slower (first run only) |
| Stop button behaviour | Immediate process kill | Stops after current day finishes |
| Recommended for | Users who already have Python | Anyone |
GNU General Public License v3.0 — provided as-is.
- Not an official Garmin product: This tool is not affiliated with, endorsed, or supported by Garmin.
- Unofficial API: Garmin Local Archive uses Garmin's unofficial API — it may change or break without notice.
- Not medical advice: All health metrics, reference ranges, and dashboard data are for personal informational use only — not a substitute for medical advice.
- Context data: Weather data is provided by Open-Meteo and Brightsky (DWD), pollen data by Open-Meteo — accuracy and availability are not guaranteed.
- Early stage: Core functionality is stable. APIs and internal structure may still change.
- No guaranteed support: Development happens when time and interest allow.
- Use at your own risk: I am not responsible for data loss or Garmin account issues.
- Feedback welcome: If something feels off — logic, structure, results — open an issue.
Windows may show a security warning ("Windows protected your PC") for either version. Click More info → Run anyway. This happens because the .exe is not code-signed. The source code is open at github.com/Wewoc/Garmin_Local_Archive — you can review it before running.
Step 1 — Extract the ZIP
Download and extract. The folder must contain:
Garmin_Local_Archive.exe ← double-click to launch
scripts/ ← all .py files — must stay next to the .exe
info/ ← documentation (optional)
scripts/is required. Without it no buttons will work.
Step 2 — Install Python and dependencies
- Download Python 3.10 or newer from https://www.python.org/downloads/
- Run the installer — tick "Add Python to PATH"
- Open a terminal and run:
pip install garminconnect openpyxl keyringStep 3 — Run the app
Double-click Garmin_Local_Archive.exe.
Step 1 — Extract the ZIP
Download and extract. The folder contains:
Garmin_Local_Archive_Standalone.exe ← double-click to launch
info/ ← documentation (optional)
No scripts/ folder needed — everything is embedded inside the .exe.
Step 2 — Run the app
Double-click Garmin_Local_Archive_Standalone.exe.
The first launch may take a few seconds longer than usual. Windows Defender and other antivirus software sometimes scan self-contained executables on first run. This is normal.
The app window is divided into two areas:
Fixed top area — always visible regardless of which tab is active:
- Connection indicators (Token / Login / API Access / Data)
- Archive status (fail / recheck / missing / range / coverage / last sync)
- Device table
- Daily Actions: Daily Sync, Mirror, Timer, Documentation
Tab area — five tabs:
- Home — Dashboard viewer (HTML dashboards)
- Files — Excel viewer
- Settings — all configuration panels
- Chat — native chat panel (v1.6.6, renamed from "Ollama-Chat" in v1.7.2). Backend: local Ollama or Cloud LLM (Anthropic/OpenAI, v1.7.2). Source: archived daily summary or live MCP tool-calling (v1.7.2). Streaming replies, session history with resume
- MCP Server — local LLM access to your archive via the Model Context Protocol (v1.7)
-> For how to use each tab day-to-day, see USER_GUIDE.txt.
-> Export Scan (v1.7.7): the button "📂 Export Scan" under Settings -> Data Collection, next to "API Scan", opens a window that shows what your own Garmin export holds (the guide is its Help button: README_EXPORT_SCAN.md). Since v1.7.7.1 the button "Rows ..." at a family shows up to 12 of its fields side by side by date.
-> Don't have your own archive yet? A synthetic 180-day demo archive
(ZIP + its own README) is available from the project's GitHub repo
under src/support-tools/demo-export/ — not bundled with this app
install, download it separately, then point GARMIN_OUTPUT_DIR at it
to try the MCP Server tab without a real Garmin account.
Settings tab (third tab):
- Email — your Garmin Connect login email
- Password — your Garmin Connect password (stored securely in the Windows Credential Manager, never written to disk as plain text)
- Data folder — where to store data (e.g.
C:\Users\YourName\local_archive) - Sync mode —
recentfor daily use,rangefor a specific period,autofor full history (everything since your oldest device — can take hours, not recommended, rate limit risk, use Bulk Import instead) - Export Dashboard Range — used by Excel/dashboard export only (not the separate Export Data feature, which always uses the full archive). Leave empty for the last 30 days.
- Age / Sex — used by the Health Analysis dashboard for reference ranges
- Mirror folder — optional second location for your archive (NAS, USB, external drive). Leave empty to disable. Set once, then use the Mirror button to sync.
- Theme (v1.7.1.8) — colour scheme for the app window and HTML dashboards, chosen from the Settings → Design dropdown. Six built-in themes (Monochrome + Rust Accent — default, Violet/Legacy, Amber & Copper, Olive & Sand, Toxic, Ice Blue). Choosing a theme opens a restart-to-apply dialog — themes are not applied live to an already-open window.
- Delay min / max (s) — randomized pause between individual Garmin API requests (default: 5 / 20). Garmin Connect has no official public API — this tool uses the same endpoints the mobile app itself calls. A randomized delay within this range keeps request timing within normal usage patterns and helps avoid rate limiting (HTTP 429). Lower values increase the risk of a temporary IP ban — 5/20 is the recommended minimum.
- Auto-apply updates in Daily Sync (v1.7.2.4) — checkbox in its own "Daily Sync" section. Saves immediately, no separate Save step needed. See "Updating the app" below for what it does.
- Migrate data in Daily Sync (v1.7.6.4) — checkbox in the same section, below the one above. Saves immediately. After an update that makes the app rewrite its summary files once, the scheduled daily sync makes a checked safety copy of your archive and does the rewrite itself, instead of stopping and asking you to open the app. Without a checked copy it changes nothing. Off by default. Details in the User Guide, Section 7.
Click Save Settings — settings are remembered between sessions.
(v1.7.2.4)
When a new version is available, a popup appears on startup: Open GitHub (opens the release page to download manually) and Dismiss, same as before.
A third button, Update, also appears — for both the Standard and the Standalone version. Click it and the app downloads the new version, verifies it against a published checksum, applies it, and restarts — no manual download, no re-extracting a ZIP. If the checksum doesn't match (a corrupted or incomplete download), nothing is changed and the app keeps running on the current version.
Unattended updates via Daily Sync (opt-in): tick Auto-apply updates in Daily Sync (Settings, its own "Daily Sync" section) and the scheduled daily sync applies a new version by itself the next time one is available — no popup, nothing to click. If the app happens to be open at that moment, it closes automatically (any unsaved changes in open fields are lost) and does not reopen afterward; the daily sync itself is unattended, so nothing would open it again automatically. Off by default.
A backup of the previous version is kept (overwritten by the next update) as a safety net, and update files are verified before anything is changed either way. (v1.7.2.4.1) For the Standalone version's GUI Update button, that backup is now also used automatically: the app restarts itself after applying the update, and if the new version fails to start, it rolls itself back to the previous working version without any manual steps. The Standard version's Update button still applies the update and closes as before — start it again yourself to use the new version.
Your password is stored in the Windows Credential Manager (the same secure vault used by browsers and Windows itself). It is:
- Encrypted by Windows using your login credentials
- Never written to any file on disk
- Only readable by your Windows user account
To remove the stored password: open Windows Credential Manager → Windows Credentials → find GarminLocalArchive → delete.
All settings except the password are saved to:
C:\Users\YourName\.garmin_archive_settings.json
Delete this file to reset all settings to defaults. The password must be cleared separately via the Windows Credential Manager.
The top section shows two things at once:
Connection indicators (Token / Login / API Access / Data) — updated automatically when Sync Data runs. Green = OK, red = failed, grey = not yet tested. No manual test button — the connection is verified automatically before every sync.
Archive info panel — populated on startup from your local data, no sync required:
- Days — total days tracked in the quality log
- high / std / fail — breakdown by quality level (colour-coded).
high= intraday data present,std= full daily data without intraday (typical for older devices or older days),fail= nothing usable - Device table — one row per device showing date range, days high, days standard, total. Double-click the
unknownrow to assign a name to legacy entries (vívoactive era and similar); double-click the row of a device that is shown with its number to name that device (since v1.7.6.9 the days of a Garmin export can be assigned to the watch that recorded them: Sync Garmin asks first, see the USER_GUIDE) - Recheck — days flagged for re-download by the background timer
- Range — earliest and latest date in your archive
- Coverage — percentage of days present vs. possible days in the date range
- Last API / Last Bulk — most recent date imported via live sync or bulk import
- Source — total source files archived · files present within the last 180-day window (e.g.
175 days · 180/180d). The 180-day window is the Background Timer's own retry window — see USER_GUIDE.txt, Section 5.
The panel refreshes automatically after every Sync and Bulk Import.
-> For what the Background Timer does with this data, see USER_GUIDE.txt, Section 5.
App doesn't start
Standard: Make sure the
scripts/folder is in the same folder as the.exeand contains all required files. Standalone: Open your data folder in Windows Explorer and navigate togarmin_data\log\fail\. Open the most recent.logfile in Notepad — it contains the full error output. If the app never started and no data folder exists yet, use the Copy Last Error Log button if the app partially loaded, or re-run from the Standard version with Python to see terminal output.
Login fails — if Garmin requires MFA, the app will show a code input popup automatically. Enter the code from your Garmin app or authenticator.
First login (no saved token) — the app shows a confirmation dialog before starting the SSO login. This is intentional — garminconnect sends several requests to Garmin during login and may trigger rate limiting if repeated too quickly. An encryption key is generated automatically in the background (no password required). Click Proceed to continue or Cancel to abort.
Token expired (sync starts slowly) — If the log shows repeated 401 errors or DI token refresh failed at the start of a sync, the saved token has expired. The app recovers automatically — do not cancel. The re-login sequence takes 3–4 minutes. Wait for ✓ Login successful before assuming something is wrong.
Standalone: If login fails due to captcha or browser verification, download the Standard version, install Python, and run
garmin_collector.pyonce in a terminal to complete verification. After that the Standalone version will work normally using the saved session.
Password not saved between sessions — click Save Settings after entering your password.
Standard: If keyring is unavailable:
pip install keyring.
Antivirus flags the EXE — this is a false positive common with PyInstaller-built executables. The source code is fully open at github.com/Wewoc/Garmin_Local_Archive. You can whitelist the file in your antivirus settings or build the EXE yourself from source.
-> For sync errors, missing dashboard data, context/weather issues, or background timer questions, see USER_GUIDE.txt, Section 8.
Standard version only.
To rebuild after modifying scripts:
python build.pybuild.py will automatically install PyInstaller if missing, move scripts to scripts/ and docs to info/, build the EXE, and create a ZIP ready for distribution.
To build the Standalone version:
python build_standalone.pyNote: This section covers optional third-party tools (Ollama, Open WebUI, AnythingLLM). It is not actively maintained and may fall out of date — check each tool's own documentation for current install steps and options.
Connect a local AI model to your health data. With local models, everything runs on your machine and your data never leaves your PC. Cloud backends and Ollama cloud models are the exception — see the warning below.
⚠️ Before you start: Both the built-in chat and the prompt file used by the external options contain your personal health metrics. If you use a local model (Ollama), your data stays on your device. If you use a cloud service — including an Ollama cloud model (name ending in-cloud) — remove any identifying details before uploading — name, date of birth, account information. AI interpretations of health data can be plausible but wrong. Always verify concerning findings with a healthcare professional.
- Install Ollama: https://ollama.com/download
- Pull a model that fits your GPU (see table below)
- Open the Chat tab in the app, choose Backend Ollama, click Start
That's it — no Docker, no separate desktop app. The panel loads the same
health-analysis system prompt used by the external options below. By
default it works against the archived daily-summary snapshot; switch
Source to MCP tool-calling (v1.7.2) to let the model query your full
archive live instead — see README.md's own Chat section for the
Backend/Source matrix and streaming/session-history details. For document
upload / knowledge-base RAG across your whole archive, use Open WebUI or
AnythingLLM below instead.
Which model fits your GPU? Rule of thumb: VRAM in GB − 2 = usable model size (Q4 quantization, Ollama's default). Rough guide, not a guarantee — actual headroom depends on context length and what else is using the GPU.
| VRAM | Rule of thumb | Example |
|---|---|---|
| 8 GB | ~6B | qwen2.5:7b |
| 16 GB | ~14B | qwen3:14b |
| 24 GB | ~22B (next common tier: ~30–32B) | — |
| 48 GB | ~46B (next common tier: ~70B) | — |
No dedicated GPU / CPU-only also works — just noticeably slower per reply, the panel's elapsed-time counter is there for exactly this case.
- Install Ollama: https://ollama.com/download
- Pull a model:
ollama pull qwen2.5:14b - Install Open WebUI via Docker:
docker run -d -p 3000:8080 --gpus all \
-v open-webui:/app/backend/data \
-e OLLAMA_BASE_URL=http://host.docker.internal:11434 \
--name open-webui --restart always \
ghcr.io/open-webui/open-webui:cuda- Open http://localhost:3000 → Workspace → Knowledge → + New → point to
local_archive/garmin_data/summary - In chat: type
#→ select the knowledge base
- Download AnythingLLM Desktop: https://anythingllm.com
- Connect Ollama (Settings → LLM → Ollama)
- New Workspace → Upload documents → point to
local_archive/garmin_data/summary
| Built-in Chat | Open WebUI | AnythingLLM | |
|---|---|---|---|
| Setup effort | None (Ollama only) | Medium (Docker) | Low (desktop app) |
| Chat interface | Basic, in-app | Full-featured | Clean, focused |
| Document/RAG quality | None — summary data only | Good | Very good |
| Best for | Quick questions without leaving the app | General AI assistant + health data | Primarily health data Q&A |
Tip: upload garmin_analysis.json directly into a chat for targeted analysis — it contains pre-processed comparisons against your personal baseline and reference ranges.
Example questions:
- "How was my sleep and HRV last week?"
- "Which days had Body Battery below 30?"
- "Compare my resting heart rate this month vs last month."
- "Based on the analysis file, which metrics need attention and why?"
Note: Same caveat as above — MCP client support in third-party tools is evolving quickly. Check Open WebUI's own documentation if this no longer matches what you see.
Start the MCP Server first (MCP Server tab in the app, or the standalone mcp_server.exe) — see USER_GUIDE.txt, Section 11, for backend, port, and Docker-specific settings.
In Open WebUI: Settings → Connections → MCP Servers → Add Connection.
- Type: MCP Streamable HTTP
- Name: any label, e.g.
GLA MCP - URL:
http://127.0.0.1:<port>/mcp— if Open WebUI runs inside Docker, usehttp://host.docker.internal:<port>/mcpinstead (the default port is8756; enable Extra allowed hosts on the MCP Server tab first, see USER_GUIDE.txt, Section 11) - Authentication: None
Save the connection, then enable it in a chat via the tools/connectors picker.
Claude Desktop and other MCP clients connect the same server differently — usually via a JSON config file (mcpServers entry) rather than an in-app dialog. Check the client's own MCP documentation for the exact syntax; the server URL and port above stay the same regardless of client.
MCP tool-calling puts different demands on a model than the built-in
chat — the model must reliably fill required tool arguments (e.g.
date_from/date_to) and pick the right field name. Based on
internal tests across multiple sessions (v1.7.1.11–v1.7.1.15, over
1,300 tool calls across 9 local models):
| Model | Suitability | Notes |
|---|---|---|
qwen3:4b and larger (8b, 14b) |
Recommended | Most reliable field resolution and tool use across all tests; larger variants noticeably slower per reply |
qwen2.5-coder:7b and larger |
Usable | Works reliably, but occasionally gets stuck repeating a call on ambiguous or empty results instead of answering |
qwen3:0.6b |
Limited | Noticeably weaker than the rest of its own model family — sometimes misreads server error messages |
command-r7b, hermes3 |
Not recommended | command-r7b consistently refused all tool use in testing; hermes3 picked the correct field in only a fraction of cases |
gemma (any size) |
Not usable | No native tool-calling template in Ollama — cannot call MCP tools at all |
Rough guide, not a guarantee — results depend on the exact question phrasing and may shift with future Ollama/model versions. Test tool and methodology: mcp-llm-tester.
Two ad-hoc timing runs give a rough sense of per-model reply speed —
tendency only, not a controlled benchmark. Raw per-question timings in
mcp_test/results/ (not published as part of this repo).
Run 20, 2026-09-14 — 25 simple single-field questions per model:
| Model | Avg | Median | Max | Timeouts/errors |
|---|---|---|---|---|
granite4.1:3b |
10.2s | 6.6s | 38.6s | 0/25 |
granite4.1:8b |
10.8s | 7.9s | 31.6s | 0/25 |
qwen3:4b |
33.4s | 31.4s | 56.6s | 0/25 |
Run 22-2, 2026-09-17 — 60 multi-step long-range/correlation questions per model:
| Model | Avg | Median | Max | Timeouts/errors |
|---|---|---|---|---|
granite4.1:8b |
44.3s | 40.5s | 146.7s | 6/60 |
qwen2.5-coder:14b |
32.5s | 15.6s | 262.2s | 7/60 |
qwen3:14b |
104.2s | 94.2s | 218.9s | 1/60 |
qwen3:8b |
85.9s | 70.2s | 306.0s | 6/60 |
Gaps this leaves open — why this is a tendency, not a measured truth:
- Different question catalogs per run (simple single-field lookups vs.
multi-step long-range correlation questions) — times are not
comparable across the two tables above.
granite4.1:8bappears in both: 10.8s on the easy catalog, 44.3s on the hard one — same model, same hardware, just a harder question set. - Single run per model/catalog combination, not repeated — no variance or confidence data, one unusually slow or fast run can shift the average shown.
- Background system load on the test machine during a run isn't controlled or logged per run.
- Timeouts/errors (up to 7 of 60 for some models) are dropped from the average rather than counted as a worst-case penalty — a model that times out often can look artificially fast here.
- No correctness/quality pass exists yet for
granite4.1(unlike the models in the Suitability table above) — timing alone says nothing about whether its answers are usable, so it stays out of that table despite competitive-looking numbers here.