Private AI for confidential work
Your documents.Your AI.Nothing leavesthe building.
A complete AI assistant on a machine in your office. No cloud, no account with us, no outbound connection.
outbound connections
no telemetry, no licence server
hardware-bound licence
verified offline, on the machine
inference engines
VEYMEER, Ollama, vLLM, TensorRT-LLM
assistant tools
and 33 workflow blocks
Everything happens inside the wall.
Your team's laptops and phones talk to one machine on your network. That machine talks to nothing else.
Laptops
Tablets
Phones
VAULT_OS
one machine
Internet, cloud AI
Your building
Laptops
Tablets
Phones
VAULT_OS
one machine
Internet, cloud AI
Your building
Install
One sealed installer builds an encrypted enclave, locked with a master password only you know.
Add a model
Use one already on the machine, import one from a drive, or ask us to deliver a pack sized to your hardware.
Connect
Vault-OS Client finds the machine on your network by itself and checks its identity before trusting it.
A whole AI workspace, not a chat box on a server.
Every tier gets all of it. The tiers differ by how many people use it, never by how good the answers are.
Ask your own documents
Contracts, case files, reports, scans. Answers arrive with the passages they came from, so a reviewer can check them in seconds.
Reads scans and images
Scanned pages, photos and diagrams are read by a vision model. Voice notes are transcribed on the machine.
Workflows and agents
Automate recurring work with 33 building blocks. Agents run in a sandbox, with human approval where you need it.
9 clearance levels
Access by role and clearance
Each person sees only the folders their role allows, enforced on the server.
Tamper-evident audit trail
Every question, answer and file access is recorded in a hash-chained ledger that you can verify independently.
Cryptographic burn switch
An authorised, irreversible erasure of all data and keys, following NIST SP 800-88.
Inference partner
veymeer.com
Powered by VEYMEER.
A high-performance inference engine, engineered from the ground up for the GPUs in workstations and on-premises servers. Developed independently by VEYMEER and integrated into Vault-OS under partnership, it runs entirely on your machine with no network access.
- Up to 12×
- faster code editing and repeated generation, with output identical to standard decoding
- Up to 2×
- faster than Ollama on new text, on the same hardware and model files
- 100%
- identical answers when the same question is asked again
- Native
- tool calling and structured JSON output
Generation speed
Tokens per second, higher is faster
Qwen 2.5 7B
Llama 3.1 8B
Mistral 7B
Tokens per second, long-form generation. Same Q4 model files on an NVIDIA RTX 3050 laptop GPU. Benchmarked by IronGap inside Vault-OS, October 2026.
Up to 12×: VEYMEER benchmark, code editing with speculative decoding against standard decoding, Qwen 2.5 7B on an NVIDIA RTX 3050 laptop GPU.
Security that holds up to scrutiny.
The questions security teams ask first, answered in terms of how the product actually behaves.
The enclave is never exposed as a drive
The encrypted volume is attached to an access-controlled folder, never a drive letter. Other accounts on the same machine cannot browse it, even while it is unlocked.
Installation survives a power cut
If power is lost while Vault-OS is unpacking into the encrypted volume, the next run resumes where it stopped and re-verifies every file it had not finished.
Every shutdown is verified
Vault-OS confirms the enclave is locked before it reports a clean shutdown, and will not start again until any volume left unlocked has been sealed.
Every connection is verified
The machine shows its certificate fingerprint on its own unlock screen. Vault-OS Client displays the same fingerprint before it connects, so your team confirms the machine before trusting it.
Zero knowledge, by design
Only you hold the master password. IronGap keeps no copy and no key to your enclave, so no one, including us, can access your data or recover it without you.
Your hardware, or a delivered appliance.
Enterprise
Vault-OS Appliance
The bare-metal edition. Its own operating system with signed boot and a TPM-sealed enclave, and Vault-OS as the only application. GPU drivers are sealed into the image, so the delivered machine never fetches anything.
- Signed boot
- TPM-sealed enclave
- NVIDIA, AMD and Intel drivers built in
- Delivered installed, on validated hardware
PRO, MAX, Enterprise
On your hardware
Install on a Windows server or workstation you already own, with TPM 2.0 and an NVIDIA GPU. Available today.
Any tier
Sized for you
We specify a machine for your team and your models. Buy it from your usual supplier, or we can source it for you.
Built for firms that handle confidential information.
Law and IP firms
Case files, disclosures and filings that cannot touch a third-party server.
For example, you could ask
- Which of our disclosures cite prior art from before 2019?
- Summarise every clause in this contract that limits liability.
The 30-day pilot
Prove it on your own documents before you commit.
You pay the first month of the tier. We agree what success looks like before we start. If it is not met, the fee comes back in full.
Scoping call
A 30-minute call to choose the documents, the team and the questions the pilot must answer. We agree the success criteria in writing.
Install
On your hardware, or on a machine we specify for you. One sealed installer, with nothing that connects outside your network.
30 days of real use
Your team works with your own documents. We check in weekly and resolve anything that gets in the way.
Decide
If the criteria are met, the fee becomes your first month and the licence continues. If not, the fee is refunded in full.
Three tiers. Flat prices. No per-token fees.
What buyers ask first.
- Does anything ever leave the machine?
- No. Vault-OS makes no outbound connections: no telemetry, no licence check, no update ping. Licences are files verified offline against the machine’s TPM; updates arrive as signed packages you copy in.
- Which AI models does it run?
- Open-weight models of your choice, including Qwen, Llama, Mistral and Gemma. The installer includes Vault-OS, its database, the inference engines and Vault-OS Client. It does not include AI models, so you stay in control of which ones run. Import models already on the machine or from a drive, or we can supply a pack sized to your hardware at no charge.
- What hardware do I need?
- For a team of 10 to 25 people: a Windows Server 2019+ or Windows 11 Pro machine with TPM 2.0, 32 to 64 GB of RAM and an NVIDIA GPU with 16 to 24 GB of memory. We confirm the specification with you during the pilot call.
- Can IronGap see our data or recover our password?
- No, on both counts. The enclave is encrypted with a master password you set and we never hold. That also means a lost password cannot be recovered by anyone, including us.
- How does the pilot work?
- 30 days on your own documents. The fee is the first month of the tier you pilot, refunded in full if the success criteria we agree up front are not met, and credited if you continue.
- Who is behind it?
- IronGap Technologies, founded by M. Taha Halakooei, who designed and built Vault-OS. You work directly with the engineering team behind the product, not a reseller.
Talk to the engineer who built it.
A 30-minute call with the engineer who built Vault-OS. We review your documents and hardware and tell you plainly whether it is the right fit.