Download it, double-click it, work.
There is no installer to configure, no account to create, no card to enter, and no server to rent. FoundrySuite is a folder with a launcher in it, and it runs on the computer it is sitting on.
One download covers Windows and Linux; you pick the file for the machine you are installing on. macOS runs the same program from the source.
Installing takes one decision
Which is whether you want the language model to run on your own hardware. Everything else is a double-click.
Windows
Double-click install.bat once. Then double-click Launch FoundrySuite.bat for every session after that.
macOS
From the source, not the download. The bundle carries a Windows and a Linux launcher and nothing for macOS, so take the source tree and double-click Launch FoundrySuite.command. It installs itself the first time you open it and starts immediately after.
Linux
Run ./install.sh once, then ./Launch FoundrySuite.sh. Nothing is installed system-wide.
The installer creates a private environment inside the FoundrySuite folder
Nothing is written to your system directories, and nothing is added to your PATH. Delete the FoundrySuite folder and the software is gone. That also means it can live on an external drive, or on a machine you do not administer.
What it asks the machine for
Not eight gigabytes of anything, and not a single thing you have to install first. Here is the whole list, because "it runs locally" is not a free claim and you are entitled to know what it costs.
| Anything installed | Nothing. The download contains its own runtime, so there is no Python and no package manager to set up. On Windows there is no command line anywhere in this; on Linux you run one script, ./install.sh, from a terminal. On macOS you run the source and there is nothing to install first. |
| Disk | Room for the application, plus several gigabytes if you supply a model to run on this machine. |
| Memory | 8 GB minimum, 16 GB for comfortable inference |
| Graphics card | Not required. A local model runs on the CPU. |
| Internet | Only to fetch a model whose address the manifest publishes, and only if you ask FoundrySuite to research a company online. |
| Account | None. The first user you create lives on that machine and nowhere else. |
What it will not ask for
- An email address that reaches anyone else. The first account is created locally and used to sign in.
- A subscription. There is no recurring charge to keep using what you have installed.
- Your data as a condition of using it. Upload nothing, and nothing is taken.
- Permission to phone home for a license check. The app does not check in with anyone — the license is a local file it verifies against a key built into the program.
- Telemetry, analytics or a crash reporter. There is no code in the product that sends usage data anywhere.
If something goes wrong
You will not need this to install or run FoundrySuite. If it ever fails to start, everything it did on the way there is written to a plain-text log in a logs folder inside your data folder — the same folder the database lives in. Attach that file to an email and whoever you bought it from can read exactly what happened. There is also a foundrysuite doctor command for anyone who wants the full picture of what is installed and how the app resolved its own settings; nothing above depends on running it.
Where the model comes from, and what it costs you
FoundrySuite's AI runs on a language model. We do not put one in the installer, and we do not host the weights — but we do give it an address to fetch them from.
The installer carries a manifest naming the model files FoundrySuite looks for and, for each one, a pinned Hugging Face revision and a SHA-256 digest of those exact bytes. When you ask for them, FoundrySuite fetches each file and checks it against the digest before accepting it; one that does not match is deleted rather than used.
That is about 5.7 GB, once — and nothing is fetched until you agree to it. On your first run the program tells you the total, names who publishes each file, and waits for you to tick a box before a single byte moves. Say no and it records nothing, asks again next time, and leaves the download available under Settings → Model files; the bookkeeping, the contacts, the pipeline and the invoices carry on, and the application says which features are waiting instead of showing you an empty result that looks like an answer. Say yes and it runs in the background, with the AI features switching on as each file lands.
Or take a model off the list entirely. The box is drawn per model, not per file: the language model on one side, the OCR pair on the other, because the OCR projector does nothing without the OCR model and neither can read a page without the other. Untick either one and FoundrySuite stops offering it, stops counting it in the total, and will not fetch it. You can untick both, point the program at a model file of your own, or do neither and take the defaults. Unlike saying no, this is remembered — but as a setting rather than a mark against you. It lives in model-preferences.json in your own data directory, is never transmitted, and can be changed back at any time under Settings → Model files or with leadflow models --opt-out none.
One of the three files is a third party's. The language model is IBM's own. The two OCR files come from a community quantization, because DataLab publishes no GGUF of the model they trained. The digest proves the bytes are the approved bytes; it does not make that account the vendor. The Privacy Notice names every source →
Also your choice: run it somewhere else
You can point FoundrySuite at a remote OpenAI-compatible endpoint — a server you run yourself, or a commercial API — instead of a local model. It will say so on screen whenever it does, because once prompts leave the machine the privacy story changes and you need to know it.
If something is wrong
Most problems with software that runs on your own machine are a missing optional component, and the fix is to know which one.
A feature says it is unavailable.
FoundrySuite starts without its optional components — semantic search and the scanned-document reader are the two — and tells you which feature is unavailable rather than pretending it works. Everything else runs.
Inference is slow.
It is a language model running on a CPU, and speed depends on the model you chose and the machine you chose it for. A smaller model is dramatically faster. Ask before you buy and we will tell you what your hardware will actually manage.
I want staff on other machines to use it.
That is supported, and it is the one deployment where decisions are yours to make rather than defaults. It requires HTTPS, and if it is not configured FoundrySuite refuses to start rather than serving plaintext. The deployment guide in the download explains it, including the parts that are your responsibility.
How do I back it up?
Copy the data/ folder beside the app. It holds the database, your files and the logs — that folder is the backup. There is also foundrysuite backup for a consistent copy on demand.
Can I get my data out?
It is SQLite. It is a file on your disk that you can open with any tool that reads SQLite. Nothing about the export is gated behind us, because a vendor who holds your data and the only key to it is not a vendor you can leave.
What is the support situation?
You get the software, the documentation and the compliance documents in the download. For anything else, the support page says what is answered, by whom, and what we have not decided.
Try it before you pay for it
Thirty days, the whole product, no card. Install it, put your own data in it, and decide afterwards whether it earns its place.