~/privacy
Privacy
Last updated
Soup trains models on your own computer, so most of what you do with it never reaches us. This page lists what does leave your computer, where it goes and why, for three things: this website, the Soup CLI and the Soup Zero desktop app. MePlay, Inc. is responsible for it.
This website
The site has no user accounts and no forms. When you open a page:
- Our host, Vercel, receives the request, including your IP address and browser type, and keeps server logs.
- A script from Ahrefs Web Analytics counts page views, so we know which pages are read. Ahrefs' policy says how it does that.
- A Google Ads tag loads on every page, so we can tell whether an ad led to an install. It tells Google your IP address and the page you opened, and it records a click on an install or Get Started button. If you are in the European Economic Area, the United Kingdom or Switzerland, or your browser's time zone is there, it sets no advertising cookies unless you accept them, in the bar at the bottom of the page or in the control below; until then Google still receives the request, without cookies. Elsewhere it can set cookies without asking. How Google uses this. A browser setting or extension that blocks ad tags stops it everywhere, and the site works the same without it.
- Our unlisted /hello page loads contributors' profile pictures from GitHub. GitHub sees your IP address and browser type, as it would from any site that shows its content. The live numbers on the site (stars, downloads) are fetched by our server; your browser asks only us for them.
Your choice is read by a script; with JavaScript off the Google tag does not load at all.
Accepting lets the tag set advertising cookies and lets Google use the visit to measure our ads. We keep ad personalisation switched off for anyone who has made a choice here. Declining later stops new advertising cookies; it does not remove ones already set, which you can delete in your browser's settings.
The site itself sets no cookies. If you make a choice about the ad tag, it is remembered in your browser's local storage under the name soup-ads-consent. The site also stores the colour theme you choose (light or dark) in your browser's local storage under the name soup-theme; it is used only to show the site in that theme and is not sent anywhere.
The Soup CLI
The CLI runs on your computer and sends us nothing by default. It has an opt-in telemetry setting (SOUP_TELEMETRY=1); in the current release it sends nothing even when set. The CLI downloads what you ask it to download, such as models from Hugging Face or packages from PyPI, and those services see the request as they would from any program. The details are in the documentation.
The Soup Zero app
Soup Zero runs on your computer. Your data, your projects, your conversations with a model and the models you train are stored there, not with us. The app sends us no analytics, no crash reports and no logs; the only thing it sends to us by itself is the update check below. The 14-day trial needs no account, no email and no card. Its start date is recorded on your computer, and nothing about the trial is sent to the license service or to us.
By itself, the app contacts two services:
- Our website, for updates. Each time it starts, the app asks trysoup.dev whether a newer version exists. The request carries no version number, system information or id of its own; our host sees it like any web request, including your IP address. The comparison with your version happens on your computer.
- Lemon Squeezy, for the license. Activating a key sends the key and a name for this installation, which is your computer's name. After that the app checks the key about once a day while it is open, sending the key and an id for this installation. Nothing is sent during the trial.
When you ask it to do something, it contacts the service that thing needs:
- Installing the app downloads Microsoft's WebView2 runtime, only if Windows does not already have it.
- Installing the engine downloads Python, PyTorch and the training tools from GitHub, pytorch.org, PyPI and the file servers they redirect to.
- Installing an update downloads the new installer, after you click Install and restart.
- Training, verdicts, exports and Improve runs download from Hugging Face any model file that is not already on your computer. The app switches off the usage reporting of the tools it runs. One thing it cannot remove: those downloads carry the versions of Hugging Face's software, Python and PyTorch, and a session id that changes each time a tool starts. The first export also downloads a conversion tool from GitHub.
- Ask Soup, only if you saved an Anthropic or OpenAI key. Your messages in the chat go to the provider you chose, Anthropic or OpenAI, with a description of the project that is open. The full list is under What Ask Soup sends below. With a model in Ollama on this computer, none of this leaves your computer.
- Ollama runs on your own computer by default. If you point the app at Ollama on another machine, deployed models and Ask Soup messages go to that machine. That is your choice and your machine.
What Ask Soup sends
The app shows the three lists below in Settings and in its setup, and this page prints them word for word, so “this chat” and “this project” mean the chat and the project open in the app.
Always sent
- your messages in this chat
- a description of the project that is open:
- its name
- the full text of its soup.yaml
- a summary of its data (row counts, column names and problems found, not the rows)
- its latest job, verdict and deployment
- how many test questions it has
- which model is live and which versions are waiting
- the record of its last GGUF export (when it was made, how big it is, its base model and training)
- your graphics card's name and memory
- Your Windows user folder is shown as ~ and your workspace folder as <workspace>
Sent when you ask
- If it needs more to answer, it can also read
- up to 200 lines of this project's job logs
- this project's training runs
- the result of a data check
- the scores and reasons of this project's verdicts
- the versions Improve made, what it decided about each with the scores behind it and how its last run went
- the scores of its base-model comparisons
- a training's checkpoints with their steps and losses
- how many answers got better or worse in a verdict or an Improve run
- how many answers your app recorded, by day, by source and by kind of feedback
- how many corrections and good answers Improve would learn from, and how many edits it would keep aside to check the new version
- the free space on your disk and what the project's checkpoints and loop folders use
- the engine's version and PyTorch build
- whether a job is running on this computer (its kind and id when it is this project's job)
- how your app is connected (whether recording is on, the model name and this computer's local address)
- whether Ollama is installed and running, with its address, its version, this project's models in it and how many models it holds
- the engine's ready-made recipes
- its advice on which training method fits your data
- the memory a training would need on your graphics card
- what it reads is sent too, so a log line or an error message can quote a few words of your data
- When you ask about your projects it also sends their names, and nothing else about the other ones.
- When you confirm a question to your deployed model, the model's answer is sent to the provider too, so Ask Soup can read it; the card says so.
- When it prepares a card for you, what became of the card is sent too: its title, whether you confirmed it, and what changed, which job started or why it failed.
Never sent
- The questions and answers behind a verdict and the answers your app recorded are only counted: they are not sent, and Ask Soup tells you which screen to open to read them.
- API keys and tokens are removed before anything is sent
- nothing from your other projects is included
All of this goes to the provider you choose: api.anthropic.com for Anthropic, api.openai.com for OpenAI, or the Ollama address in Settings (your own computer unless you changed it).
A key you paste into the chat is replaced with a placeholder before it is sent. Anthropic's own client also sends your operating system, processor type and its library versions. Soup Zero sends each request to OpenAI with OpenAI's response-storage option switched off; what OpenAI keeps beyond that is governed by OpenAI's own data policy. The dock shows which provider your messages go to.
On your computer, the app keeps files in its own folder, in your workspace, in the engine's folder under your user profile (run history and a local record of commands with your Windows user name and computer name), in the Hugging Face download cache, in its browser component's data folder and in Ollama's model store. Your license key, and your Anthropic key, OpenAI key and Hugging Face token if you saved them, are kept in Windows Credential Manager; if Windows refuses to store the license key, it stays in the app's settings file until a later start can move it. Uninstalling the app does not remove all of these; the app's guide lists each place.
The diagnostics file is made only when you click the button. It is written to your disk and goes nowhere unless you send it to us. Your Windows user name, your computer's name, your license key and other keys are masked in it. Project names, file names and text quoted in logs are not, so open it and read the logs before you send it.
When you pay
Subscriptions to Soup Zero will be sold by Lemon Squeezy, the seller of record. It collects your email, billing address and payment details and handles tax. We receive the order record it gives a seller, which includes your name and email, what you bought and when. We never see your card number.
Donations on the support page are taken by Stripe on behalf of MePlay, Inc. Stripe collects the payment details; we never see the card number.
When you write to us
If you email us, we keep the message and your address so we can answer and follow up. We do not add you to a mailing list.
What we do with it, and what we do not
- We use order and license data to give you the product, answer support questions and keep our accounting records.
- We use page-view counts to decide what to improve on the site.
- We do not sell personal data, and we do not use your data, your projects or your models to train anything. We do not receive them unless you send them to us, for example in a diagnostics file.
- The only sharing for advertising is the Google Ads tag described above, which measures whether our own ads work.
Your choices
Write to us to ask what we hold about you, to correct it or to have it deleted. We keep order records for as long as tax law requires and delete the rest on request. If you live in a country whose law gives you further rights over your data, those rights apply and we will honour them.
We will change this page when what we collect changes, and the date at the top will say when.
Contact
MePlay, Inc. 1111b South Governors Avenue, STE 40549, Dover, DE 19904, United States. Write to team@trysoup.dev. A person on the team reads what we receive.