Your AI.Your rack.Your rules.
An on-premises private AI system with local models, governed document search, and an operating handoff your IT team can own.
What it is
Private AI in your building.
Keep is an on-premises AI system running a top open-weight model on hardware you own. Nothing your team types, uploads, or generates leaves your network.
Hardware you own
Dedicated server or workstation hardware, specified for your workload and budget, quoted at order. Sourced and configured with known provenance and manufacturer warranty.
Configured to your team
Concurrency, context length, and storage are set from what your team needs to run, and the server is sized to match.
The model you choose
Llama, Qwen, Mistral, DeepSeek, or your own weights, chosen for your workload and budget. You can change models later.
Works unplugged
The model, inference engine, and document retrieval run locally. The system keeps working with the network cable unplugged.
Who it is for
For work that cannot leave.
Sending documents to a hosted model means someone else holds them, prices your access, and can change the terms. A machine you own does not.
Legal & privileged counsel
Client files stay inside the firm
Client files, litigation records, and depositions covered by privilege should not reach a third-party model. With Keep, you can say they never left the building.
Healthcare & clinical operations
Patient records stay on site
Patient charts and clinical notes stay in your building. No outside model vendor receives them, so there is no model vendor to sign a business associate agreement with.
Government contracting
Controlled unclassified information
Capture, proposal development, technical analysis, and contract review inside an enclave you govern, alongside the CUI controls you already run.
Executive & proprietary IP
Board materials and trade secrets
M&A strategy, financial models, board minutes, and engineering records stay on a machine in your own locked IT space, not on a vendor's servers.
How it works
One network. No outbound path.
The system connects to your internal network over private HTTPS. Remote staff connect through your existing VPN or zero-trust access path. There is no public inbound port and no outbound path for data.
Authorized users
Your network
Customer-controlled perimeter
No public inbound port- Open-weight model runtime
- Dedicated GPUs
- Document search with citations
No telemetry · runs offline
The acceptance test
Proof before you accept.
Before you accept the system, we run your workload on it, on your premises, in front of you.
If the system does not pass on the agreed hardware, we fix it or you return it and we refund everything but shipping. That is the whole promise.
Simultaneous team load
Your team generates simultaneously under agreed peak concurrency for your build, each with large documents loaded into context.
Agreed speed per user
Every concurrent stream maintains agreed interactive tokens per second individually under full load.
Unplug the network cable
We pull the physical network cable and repeat step 1. The system continues generating without pause.
Inspect the machine logs
We show you the machine's own logs in front of your IT team: zero bytes left the building.
Specifications
Specifications.
Final figures are set for each build and confirmed at acceptance.
Compute
- Model runtime
- Any open-weight model, evaluated and selected for your workload and budget.
- Hardware
- Enterprise GPU accelerators
- System memory
- Storage
- High-speed NVMe
Performance
- Concurrent capacity
- Configured from your request and the server specification.
- Context window
- Configured to your workload and document corpus.
- Stream throughput
- Sustained interactive speed per stream under full concurrency.
On your premises
- Physical form
- Quiet workstation tower
- Rackmount chassis
- Network
- 10GbE
- 2.5GbE
- Power
- Tower · 120V / 15A standard circuit
- Rack · datacenter feeds
- Noise
- Office-quiet towers, verified on-site at acceptance.
Proof and terms
- Data egress
0 bytes outboundConfirmed by local logs and packet capture.- Delivery
- Four weeks from approved specification to a working system on your floor.
- Acceptance remedy
- Full refund except shipping if the acceptance test does not pass.
- Tower, rack, and multi-node tiers
- 5 to 500+ named users
Continuity
What you keep without us.
Skilak is a young company. The system is built so it does not depend on us.
The software is open source
Nothing in the stack is ours to withdraw, and none of it phones a license server. A system that stops hearing from us keeps serving.
There is no vendor backdoor
No standing remote access, no permanent tunnel, and no unmanaged support account. Support runs through a path you approve and can revoke. Removing us is a firewall rule, not a migration.
The runbooks are yours
Install, backup and restore, egress verification, and network and access design are written procedures handed over at acceptance, not knowledge held in one person's head.
The hardware is bought, not leased
You own the machine outright. There is no per-seat AI fee to lapse and no subscription that bricks the box.
Frequently asked questions
Questions teams ask first.
Where it lives, who can reach it, and what happens if it fails the test.
Where does the machine physically live?
In your building—in an office, a locked server room, or a secure network closet. You own the hardware outright.
How do users connect to the server?
Over your local network via private HTTPS. Each user has their own session. Remote staff connect through your existing corporate VPN or zero-trust access.
Does it keep working with the internet unplugged?
Yes. Model weights, the inference engine, and document retrieval all run on the local hardware. Unplugging the network cable during generation is step three of our acceptance test.
What happens when a newer open-weight model is released?
You can switch. Keep runs open-weight models, so when a better one is released we help you evaluate it and move to it if it is worth it.
What if the system fails the acceptance test?
If the system does not produce the agreed token throughput for your configured concurrent workload, continue working with the cable unplugged, and prove zero egress in the logs, we fix it or you return the hardware and we refund everything but shipping. That is our entire guarantee.
Who has access to the server once installed?
Only the people you authorize. There is no vendor back door, no telemetry phone-home, no secret administrative account, and no persistent external tunnel. Removing Skilak is a firewall rule, not a migration.
Does buying Skilak Keep make us compliant?
No. No product creates compliance by itself. Skilak Keep is designed to operate inside your existing controls, policies, and evidence requirements; certification and accountability remain with your security program.
Pilot
Start with one workload.
Tell us who uses it, what it can read, and what must stay inside. We will map the build.
