Private AI on hardware you own, sized to your company.
Kodros sets up a server you own, pairs your computers and phones to it, and keeps a backup we have tested. Local AI runs on that server, one approved job at a time.
What Kodros does
-
A server you own
An always-on computer in your office holds your shared files and records. You own the hardware, and the accounts open in your name.
-
Backups you have tested
Three copies: the live data, a second copy on separate storage, and one kept offline or locked against change. We restore a real file while you watch, before you sign off.
-
Local AI on your server
The model runs on your hardware, so your questions and files are read there. A job goes outside only for reasons you allow, and you can allow none.
-
Connectors, read-only first
Email, calendars and shared drives link in one at a time. Each starts read-only, gets the least access the job needs, and can be switched off.
-
A named person approves
Sends, payments, posts and access changes wait for a yes from someone you name. If nobody answers, nothing goes out. An agent never deletes your data.
-
Desktop and phone, paired
We install Hermes Desktop, open-source software from Nous Research, pin it to a tested version and pair it to your server. The Kodros iPhone app is a public TestFlight beta. Devices reach the server over a private network.
We start with one job you choose, such as drafting replies to a common customer question. It runs read-only while you watch, then you decide. If you leave, you keep the hardware, the files and the memory.
Sized to your company
Memory sets the size of a local install. The model has to fit in GPU memory, with room for each person using it at that moment.
| Staff | Concurrent users | Model class | Total VRAM | Hardware fit |
|---|---|---|---|---|
| 1-10 | 2 | 8B | 10.0 GB | One 16 GB GPU |
| 11-25 | 5 | 32B | 45.4 GB | One GB10 box, 128 GB unified memory |
| 26-50 | 10 | 32B | 70.1 GB | One GB10 box, 128 GB unified memory |
| 51-100 | 20 | 70B | 168.1 GB | Two GB10 boxes linked, 256 GB: the setup Kodros runs itself |
| 101-250 | 50 | 70B | 353.4 GB | More than two boxes, sized per job |
Assumptions: 20% of staff at once, a 16,384-token context each, 4-bit weights, FP16 KV cache, 15% runtime overhead. Each row is the top of its band, memory only.
Try your own staff count
- Concurrent users
- 2
- Model class
- 8B
- Total VRAM
- 10.0 GB
- Hardware fit
- One 16 GB GPU
Same formula and assumptions as the table.
Sizing, in plain words.
Why size by memory?
A model has to fit in GPU memory, along with the working context of everyone using it at that moment. The table adds both, then adds runtime overhead.
Can a job still leave the building?
Only for reasons you allow: the local model cannot do the job well enough, it is busy, it is down, or a named person sends it out. You choose which outside service may receive it, and you can allow none. Connectors such as email reach their own provider by nature.
What do I keep if I stop?
The hardware, the files and the memory. The accounts are opened in your name.
Each install is quoted from this sizing.
Tell us your staff count and the first job you want done. We reply with a written quote.