Pre-1.0 · accepting a small number of design partners Apply

Pre-1.0The secure computer for AI agents

Run AI agents your security team can say yes to.

You bring the agents. VOLENT keeps your data where you allow, runs every job once, and signs a record your auditor can check. Your servers or ours.

Book a demo See how it works

Nothing extra inside: no way to log in, nothing to install, nothing to break into.

Console · a VOLENT computer on our test server
The console: two agents, one of them planning, its step log written to the journal, and the approved-list activity, live from a real computer.
Real screen, not a mock-up: the VOLENT console, where you watch your agents, see every allowed and blocked call, and get the signed record. A preview from our test server, 17 October 2026.

Measured, not promised

150,000 agents. One small computer.

Every number comes from a test you can rerun.

  • 0jobs done twice after 5 power cuts.phase-7 image, median of 3 runs where run three timesSee the crash test
  • 150,000agents working at once, on every size we tested.Benchmark v1, single run, test model that answers in 5 msSee the scale test
  • 1,000 of 1,000harmful tools stopped. None escaped.phase-7 image, 0 crashedSee the sandbox test
  • 3.5 sto start up.phase-7 image, 3.486 s median of 3See the start-up test
How these were measured

Measured 2026-10-14 on the image built from tag phase-7, median of 3 runs where run three times, 2013 Intel i7-4770 under KVM.

Capacity figures use a 5 ms mock model, so they measure VOLENT, not a model. Certified three-run results on the final image will follow with our paper.

One computeron your servers · illustrationoperating system built only for agents
150,000agents calling an AI model at once 1approved list for every call 0actions done twice
Illustration: each dot stands for many agents. Measured: 150,000 agents working at once with a test AI model, on every computer size we tried (single run, 2026-09-23).

How it works

Everything your security team asks for.

Not a box on an ordinary computer. The computer itself, built only for agents.

Most agent tools

Your agent
A sandboxa box on someone's server
An ordinary computer systembuilt to run anything
Their cloud

VOLENT

Your agents
An operating system built only for agentsapproved liststep logsigned recordno way in
Your own serversor dedicated servers we run for you
VOLENT replaces the ordinary computer system under your agents, so the safety is part of the computer itself.

Data only goes where you allow.

  1. Your agents send data out.
  2. Approved places get through.
  3. An unknown site is blocked, and the reason is written down.

Nothing happens twice.

  1. Each step is written down, then done.
  2. The power cuts out mid-payment.
  3. On restart, finished steps are skipped. The payment is made once.

A record no one can quietly change.

  1. The computer signs what it started and where data went.
  2. Your auditor presses Check on their own laptop.
  3. Every line matches. Nothing was changed.

See it in the real files.

1. Describe the agent. One short file per agent. Anything unexpected is refused.
apiVersion: volent/v1
kind: Agent
metadata:
  name: invoice-triage
spec:
  model:
    budget:
      max_concurrent: 4
  instructions: |
    Triage incoming invoices.
  triggers:
    - kind: webhook
2. Name where it may go. Each tool is tied to one approved place. Risky steps wait for a person.
  tools:
    - name: fetch_invoice
      kind: http
      request:
        method: GET
        url: https://erp.internal/api/invoices/{id}
      capability: erp-read
  approval:
    required_for: [tool:erp-write]
3. Check the record. Your auditor runs this on their own computer. It fails if anything changed.
$ ./volent attest --verify policy.json \
    --report record.json \
    --nonce d03748c4108b21cf

# output shortened
  ok   machine identity signature — ed25519
  ok   freshness — the report carries the nonce that was sent
  ok   TPM signature over the quote — ECDSA P-256 / SHA-256
  ok   PCR values match the signed digest

PASS: the report verifies and satisfies the policy

Check it yourself

Don't trust us. Check the record.

Press Check. Change a line. Press it again.

Tamper-proof recordsample · values illustrative
  • Started up unchangedmatches the approved version
  • Data went only to approved placesmodels.example · erp.internal
  • Every step loggedsigning each step arrives with the current release
  • Signed by this computeranyone can check the signature
A sample record with made-up values. Today the machine signs what started up and where it connected. Signing every action arrives with the current release.
A real record in the console: what started, how it was checked, the security chip's measurements, where data may go, who signed it, and the challenge it was made for.
And here is a real one: a record from our test server, with the exact commands to check it.

Who it is for

Built for teams that answer to an auditor.

One answer for each person who signs off.

Head of security

Where can our data go?

Only to places on your approved list. The rest is blocked and logged.See how CTO and platform team

How hard is it to run?

One installer on servers you already run. You keep the keys.See how Risk, compliance and audit

Can we show what it did?

A tamper-proof record your auditor checks on their own laptop.See how Business owner

How do we start?

Pick one real workflow. Our engineers work on it with you.See how

See six real workflows

Straight answers

Ready today. Coming next.

An early version (pre-1.0). The honest list.

Ready today

  • The computer itself, which checks itself every time it starts.
  • An approved list for everything that leaves.
  • A step log, so a crash repeats nothing.
  • A tamper-proof record of what started up and where it connected.
  • A locked box for the tools agents use.
  • The console to watch agents and get the record, as a preview.

Not built yet

  • Agents doing a whole job on their own: the built-in agent loop arrives with the current release.
  • Signing every action arrives with the current release.
  • Cloud installs: guides are written; tested so far on our own servers only.
  • Paid AI models (like the big chat services): not tested yet; free open models run today.
  • Hiding data from whoever runs the server: Not yet — on our roadmap.
  • Certification: There are no certifications yet.
  • Signing keys: this version uses development signing keys.

Read every answer for your security review

Your servers or ours. Same safety.

Free to start. Bring one real job and we run it with you.

Managed

We run it for you.

Dedicated servers, privately linked to your network.

For design partners
Self-hosted

You run it.

Your own servers or cloud account, with one installer and a guide.

Tested so far on our own servers
Book a demo Become a design partner

Pre-1.0 · Free for design partners through 1.0 · See pricing