Research › How GroupLab is built
What GroupLab sends from your computer or phone
On this page
GroupLab project. Researched and written with Claude. Testing and data collection by the developer. Written 2026-09-22; the data is from 2026-11-06.
No figure: A list of network calls has no picture. The one thing a reader could look at, the crash report contents, is quoted in full in the article.
What we found
GroupLab sends five things, none of them without an action from you, and none containing where you were. A test fails the build if any part of the program learns to reach the outside world on its own.
How sure we are
This describes the code as it stands and the tests that hold it there. It is not an audit by anyone else, and you are welcome to read the source instead.
The whole list#
- An update check. A plain request for a small file saying what the newest build is.
- An update download, if you accept one.
- An error report, when GroupLab hits an error: sent by itself only if you said so, offered otherwise, never if you said never.
-
A target, if you choose to send one to the project: from the page at
grouplab.org/targets/, or from GroupLab itself after you analyze it. -
A hardware survey report, only if you said yes when GroupLab asked, on the same screen as the questions above: what your machine is and how fast GroupLab ran on it, at most once a week, and sooner after you run the benchmark, whose every run goes with the version that ran it. What a report holds is listed on that screen and in
docs/SURVEY.md, and it never holds your name, a file name, a photograph or a location. It does hold a random number GroupLab made for this installation, so that your repeated runs count once: it is not taken from your device, your account or your network, and Settings, Sharing and privacy, can reset it or delete your reports. The published figures are at grouplab.org/survey/.
That is all of it. There is no analytics, no usage reporting, no license check and no phoning home. Nothing is sent while you are marking a target, and nothing is sent because you opened the program, except what you chose to have sent by itself: error reports, and the survey's weekly report.
The update check#
A GET for a small JSON file. The only thing it says about you is a User-Agent of GroupLab and its version number, because a server needs to know which build is asking in order to answer sensibly.
It does not send your machine name, your account, your screen, your Windows version or a unique identifier of any kind. There is nothing to correlate one check with another beyond the fact that somebody on that version asked.
You can turn it off, and then nothing is sent at all.
On Android, only GroupLab Dev checks, the build installed from its APK: the same GET, for the same kind of file, when it starts and about every six hours, and the download itself only on Wi-Fi unless you tap Update now. The copy from Google Play never checks; Google Play updates it. "Install updates automatically" in Settings, Updates and about turns off the installing, not the looking.
An error report, and only as you chose#
When GroupLab hits an error, or closes without shutting down, it writes a record to a folder on your own machine. The first time it can, it asks whether to send such reports to the project: automatically, only when you say so each time, or never. Until you choose, it asks each time.
What a report holds: the version of GroupLab and the system it runs on, the error and where in GroupLab it happened, and the names of the last few things done, such as setting the caliber, with nothing that was typed into them. Since 4 October 2026 on the computer, and since nightly 173 (6 October 2026) on Android and the iPhone, an automatic report also carries GroupLab's log from this run and the last, with anything typed replaced by its length; GroupLab asks once more before it does, those who chose automatic reports. Never a photograph or scan, a file name, a location, or anything you wrote. The same error several times goes as one report with a count, and a day's reports are capped at 20.
It goes to grouplab.org, not to GitHub, and the application holds no key or token. The project's server turns it into an issue in a private repository, and the token that lets it do that lives on the server and nowhere else.
A report you agreed to send is tried again. If it cannot go at that moment it is kept on your machine and tried again for seven days, never twice once it has gone, and never with a dialog in your way. A report made by hand, from the banner or Report a problem, goes once, as a zip you can look at first, with a description of up to 500 characters if you write one.
Two things about all of it are deliberate:
No authentication. There is no key, because any secret inside an open source program is public the moment it ships, and pretending otherwise would be theatre. So anybody can send a report, and nothing in one is ever taken as an instruction.
No metadata, no paths. The log carries no file paths and no image metadata. That is a rule with a test behind it rather than an intention.
Your photographs, and what is stripped#
If you send a target from the page, that is an upload you started, to a page you visited, with a consent record written at the time.
GroupLab can also send one itself, and it asks first. After you press Accept and analyze on a target, a short panel under the figures offers to send it: the image, the holes GroupLab found, the ones you moved, added or removed, what you told it, the figures and the log from that session. Nothing goes until you press Send and choose one of two levels of consent, testing only or may be published. Settings lets you say instead that every target should go, or none. The first time GroupLab opens after an update that carries it, one screen asks which you want, and nothing is chosen for you.
Since 4 October 2026 on the computer, and since nightly 173 on Android and the iPhone, a third choice: Send everything I open. Every picture you open goes when you leave it, whether GroupLab could read it or not, with what it worked out, the state it was left in and the log, because a picture GroupLab could not read is the one that shows what to fix. The old automatic choice is now Send finished targets only, under which a picture GroupLab could not read, or one left before Accept and analyze, is never sent; nobody is moved from one to the other without being asked. On a phone a finished target can be sent the same way, and whatever is sent waits for Wi-Fi unless you allow mobile data in Settings.
Unlike a crash report, a target you agreed to send is tried again. If it cannot go at that moment it is kept on your machine and tried when GroupLab next starts, for seven days, then deleted. Settings shows anything waiting, with a button to discard it. That is the one thing GroupLab may send when it opens, and only because you already said yes to it.
The photograph leaves with its pixels and without its place. GroupLab removes the location, the time, the camera's serial and everything else a phone writes beside the picture, and checks that the picture itself is untouched. Where it cannot do that without changing a pixel it saves the picture again without loss, and where even that would be too large it does not send it and tells you why.
Location is never read. Not stripped on the way out: never read in the first place, at any point, by any part of GroupLab. A photograph taken on a phone usually carries where it was taken, and a target photograph usually means somewhere you shoot. That is not information this project wants to hold, so the code does not look at it.
Sync between devices#
If you turn on Sync between devices in Settings, GroupLab writes your records to a folder you choose, and only there. It sends nothing over the network for it, so it is not a sixth item on the list. Your own cloud app, such as OneDrive, Google Drive, Dropbox or iCloud Drive, is what moves that folder, and the project never holds it. The original photographs go into the folder only if you ask.
Where it is kept, and for how long#
NOTES-FROM-PLANNING.md entries 215 to 217. Nothing people send stays on the web server longer than it is needed, and every place it can wait has a limit.
- A target, from the page or from GroupLab. Within minutes it is rebuilt from its pixels and the file you sent is deleted. The rebuilt copy waits on the server until the server's own archive worker, or the developer's pull, copies it into the project's private archive hosted by GitHub, proves the archive holds it, and then deletes it from the server; the pull also keeps a copy on the developer's machine. One the server refuses is deleted after seven days; one nobody pulls is deleted after sixty.
- Once kept by the project, a target is on the developer's machine and in the private archive, which is never made public. It is published only if you chose "may be published", and only after it has been looked at. The archive keeps to a budget, and the oldest submissions in it may be deleted to stay inside it, never while one is still used by a test or an article.
- Read again by every build (entry 394). A target sent from GroupLab with your corrections is read again by each new nightly, in the private archive's own automatic job, and compared with what you kept, moved, added and removed. Only numbers are kept: how many holes were found, missed or invented, and how far off they were. Never the picture, its location or anything that names you, and the numbers of a target sent for testing only never appear in anything public.
- An error report becomes an issue in the project's private repository on GitHub, and the copy on the server is deleted the moment the issue is opened or updated. One that cannot be sent is deleted after thirty days.
- On your own machine, a target you agreed to send and could not is tried again for seven days, then deleted, as above.
- A hardware survey report is counted into totals within the hour and deleted. None is kept on the server longer than thirty days whatever happens. What stays is counts: for each installation, only a keyed hash of its random number (the key exists only on the server, so the stored value cannot be matched to anything else), the classes its machine falls in, the month it was last seen, and for each version of GroupLab how many benchmark runs it sent and how their times fall, from which its median is read. The runs themselves are not kept. A machine not seen for twelve months is dropped. Delete my survey reports, in Settings, removes everything kept under your number at the next hourly count; that is possible only because the number stays the same, which is why it is a number GroupLab made rather than nothing at all. Nothing smaller than ten machines is ever published.
The rule that stops a sixth thing appearing#
All five go through one interface. Nothing else in the program is allowed to open a browser, fetch a file, start a process or read the clipboard.
That is not a convention, it is a test. OneWayOutTests reads every source file in the repository and fails the build if anything outside that one file starts a process with the shell or launches a URI, and a second guard does the same for the clipboard.
It exists because of a real accident. The settings page's link to the repository once opened the real default browser, and an automated test that clicks every control it finds opened browser tabs on the developer's machine. The same class of mistake had already printed to a real printer. Both were fixed the same way: one door, and the tests replace it with a recorder that writes down what would have happened.
That is the part worth trusting more than any promise on this page. A rule a person has to remember is a rule that lasts until somebody is in a hurry. A rule that fails the build is a different kind of thing.
Reading the clipboard#
GroupLab can open an image you paste with Ctrl+V. It reads the clipboard only at that moment, never on its own and never in the background, and the same one-door rule covers it: exactly one file in the program may touch a real clipboard, and the build fails the day a second one learns how.
Choosing a photograph on the phone#
GroupLab for Android asks for no permission to your storage or your photo library. Choose a photograph opens Android's own photo picker, and From another app opens the app you pick, such as Google Photos or Samsung Gallery; either one hands GroupLab only the picture you chose, for as long as it takes to read it. A picture shared into GroupLab from another app arrives the same way. GroupLab asks the app for the picture's size and nothing else about it, never reads where it was taken, and keeps only the working copy it measures, which carries no metadata at all.
What this means#
You can read this list and then check it. The point of the test that fails the build is that the list stays five items long without anybody remembering to keep it that way. If you would rather verify than trust, the test is named in the sources and it is short.
Nothing here is a promise about your network. It is a statement about what this program does. If something on your machine is watching what it talks to, this article tells you what it should see and nothing more.
The part worth carrying away: a program that reaches the outside world in five places can be described in a paragraph, and one that reaches it in forty cannot be described at all. That is the reason for the rule, rather than any particular one of the five.
Sources#
- The one way out, and the test that keeps it the only one:
src/GroupLab.Core/Updates/OutsideWorld.csandtests/GroupLab.Core.Tests/OneWayOutTests.cs. - What a crash report contains and how it is sent:
docs/CRASH-REPORTING.md.