An agent runtime that installs inside the company. Your content never passes through our servers: it goes from your machine to your AI provider, or it does not leave at all if you use a local model.
Nothing runs without your yes
Audit chain
On screen
Real screenshots of the application
What it is
European companies bound by professional secrecy or sector regulation — law firms, accountants, engineering practices, factories — hold exactly the documents an AI agent would get the most out of, and exactly the ones they cannot upload to someone else's cloud. iForja solves this from the other end: instead of uploading the documents to someone else's platform, it takes the runtime to the client's server, with their own LLM key, every permission denied by default, and a sealed audit chain of everything the agent does. What the agent sends to the model goes straight to the provider the client chooses; with a local model, nothing leaves.
Who it's for
European companies bound by professional secrecy or sector regulation — law firms, accountancies, engineering practices, factories under signed confidentiality agreements — and which therefore cannot take their documents to someone else's cloud.
How it works
Three ways to have it, none of them rented
You download it and it runs on your own machine, connected to your own model key. Or we install it on your server, with your GPU, and everything runs in there. Or we supply the server with its GPU, from a DGX Spark upwards. In all three the system is yours and it sits where you are.
The model key is yours
iForja brings no model of its own: the client provides their key, stored encrypted with AES-GCM. Seven supported providers, or a local endpoint. The company chooses and pays for the compute, and iForja takes no cut of tokens: changing model does not mean changing system.
No action without prior approval
Every action with effects — shell, files, web, an MCP server, activating a skill — passes through a deny-by-default approval gate, and the operator approves or rejects it from the console. Sixty-one role-based access rules decide who may approve what. And if they cannot be resolved, the agent closes down to zero tools instead of pressing on.
Every action is sealed
Each event is chained to the previous one with HMAC-SHA256. Deleting or editing an entry breaks the chain and the verifier flags it: you do not have to trust the log, you check it. It lets you answer a client, a partner or an inspector about what was done with their data, without asking them to take your word.
Nothing to repatriate the day you stop
Because your documents never passed through our servers, stopping does not mean recovering anything from anywhere or negotiating an export. The system shuts down and the information stays where it always was.
Frequently asked
Why not use a cloud service that already promises not to train on my data?
Because a contractual promise not to train on your data is not the same as your data never leaving. With iForja your content never passes through our servers; with a local model it never leaves your premises, and with a cloud provider's key it goes only to that provider, which you choose. For a firm bound by professional secrecy, or a company that has signed confidentiality with its own clients, the gap between “we don't use it” and “we don't have it” is the gap between having to trust and not needing to. And there is a second, more practical reason: an agent that takes actions on real documents needs explicit permissions and a record of what it did with each one. No chat gives you that, however good its privacy policy.
iForja has a site of its own: how it works, what leaves your machine and where it goes, and how to check it yourself.
iforja.comEach deployment is installed one at a time, fitted to each client's infrastructure.