Private AI, on a server you own.
An AI assistant that runs on your own server, so nothing you type or upload leaves it. We tested it before offering it: a small open model handles everyday drafting and sorting well, and gets calculations and subtle questions wrong. This page tells you both.
What you get
A private chat interface
A web page your team signs into, with one account per person, running on your server.
A small open model
An open model of about 2 billion parameters (in our test, MiniCPM5 2B), running on the server itself. No AI company receives your text.
Your data stays put
Uploads, conversations and settings are stored on your server and backed up with the rest of your data.
Set up for your tasks
Ready-made instructions for the jobs we agree on, such as a reply template or the fields to pull out of an invoice.
A short training session
We show your team what to use it for, what to check, and where it goes wrong.
A written handover
What is installed, how to add users, how to update the model, and how to switch it off.
What it does well
Always with a person reading the result before it is used.
- Drafting replies to emails and messages, for a person to edit and send
- Sorting incoming emails by topic or urgency
- Pulling data out of invoices, forms and similar documents into a table
- Summarising long documents so you know what to read first
What it does not do well
- It is not a ChatGPT replacement. A small model knows less and reasons less well than the large models behind ChatGPT or Claude.
- It makes mistakes with calculations. Totals, VAT and dates must be checked, or worked out by ordinary software.
- It struggles with nuanced questions, such as a legal or medical judgement, and can sound confident while being wrong.
- It is slower than a commercial service, because it runs on a small server rather than in a data centre.
- What it writes is a draft. Someone has to read it before it reaches a client.
When you need more
A larger model on a bigger server
Still entirely on your own server, and noticeably better at reasoning. The server costs more each month, and we size it with you before you commit.
A private interface to an AI provider
Your team uses the same private workspace, connected to a provider whose contract says it does not keep your data or train on it. The answers are much better, but this is not fully self-hosted: your text leaves your server to be processed, under that contract.
How it works
- Step 1
Name the tasks
We agree on the jobs it should do and look at real examples with you, with names removed if you prefer.
- Step 2
The server
It needs about 4 GB of memory and 2 to 4 processor cores. If you do not have a server yet, we set one up first.
- Step 3
Install and test
We install the model and the interface and try them on your examples. Usually 24 to 48 hours once the server exists.
- Step 4
Train and hand over
A short session with your team, then the handover document. We tell you plainly which of your tasks it handled well and which it did not.
What it costs
From
$750
A fixed price to install the workspace on one server, set up for the tasks we agree at the start. The server is billed separately.
See all pricesBilled to you directly
- The server, billed by the provider: roughly USD 10 to 20 a month for 4 GB of memory and 2 to 4 cores.
- Server setup and hardening, if you do not have a server yet.
- An AI provider's usage fees, only if you choose the connected option.
What is not included
- Building a custom AI application, or connecting it to your other software: that is AI integration
- Training a model on your own data
- Any guarantee that what the model writes is correct
- Keeping the server and the model updated afterwards: available on the monthly care plan
Questions, answered
Does anything leave our server?
With the local model, no. The model runs on your server and your text is processed there. The one exception is the connected option above, which sends text to an AI provider under contract, and we only set that up if you choose it.
Is it good enough to replace ChatGPT?
No, and we would rather say so now. For drafting, sorting and pulling data out of documents, a small model is useful. For calculations and questions that need judgement, it makes mistakes a large model would not. If your work needs that level, choose a larger model or the connected option.
Who is it for?
Anyone with documents they should not paste into a public AI tool: law firms, clinics, accountants, and any business holding confidential client files.
Does this make us compliant with data protection law?
Keeping the data on your own server helps, but compliance depends on how you handle data as a whole. We document where the data lives and who can reach it, which helps your own assessment. We do not give legal advice.
Related services
Server setup & hardening
The server the workspace runs on, set up properly and in your name.
AI integration
When AI has to work inside your other tools, or needs a larger model.
Monthly care plan
Keeps the server and the model updated and backed up.
Document extraction
Invoices and receipts read automatically, and the private model can be the one reading them.
Documents you cannot send to ChatGPT?
Tell us what you want the AI to do with them. We will tell you honestly whether a small private model is enough.