Self-hosted AI gateway
One gate to
every AI model.
See every AI request. Set budgets and access rules. Reduce spend—all on your infrastructure.
Change the base URL. Keep your code.
The models you use. One place to govern them.
- Llama
Measured, not projected
$15,917 became $4,163.
Same team, same models, same output.
7-person team · approximately 2 months. Results vary by workload.
The Torii panel
Your AI, no longer a black box.
See the bill. Set the rules. Know what happened. Every rule is live on the very next request.
Every request has a cost—and a person behind it.
Usage overview
Customer sample · ~2 monthsProvider spend
$4,163.94
Saved
74%
Requests
30,297
Tokens
~3.65B
An illustrative daily-spend bar chart. The last five days are highlighted.
Soft alerts or hard stops. A budget for every team.
Team budgets
Example monthly rulesA limit for each team. Checked before the request reaches a provider.
Priced, checked and logged before the request moves on.
A request, accounted for
Example audit sequenceA traceable record of what happened—and which rules were applied.
01Request priced
Claude Sonnet · ada@acme · $0.412
02Access & budget checked
Engineering · model allowed · within cap
03Audit entry recorded
Hash-linked entry · redacted request metadata
Tamper-evident ledger. Verifiable from the panel or API.
Interactive illustration, not a live account. Overview totals use the measured customer sample; activity, budgets and audit entries are examples.
Budgets
Per user, team or key. Soft or hard caps.
Access
Who gets which model, set once.
Audit
Hash-chained. Provable. EU AI Act ready.
Failover
Across regions and providers, automatically.
One gateway. Every request priced, checked and logged.
How it works
Direct access is easy.
Control takes a gateway.
Your teams already use AI. Torii sits between their tools and the providers, turning anonymous usage into something you can govern.
Make spend visible
Attribute every request to a person, team and model.
Put rules in the path
Check access and budgets before the provider is called.
Leave a verifiable trail
Write every request to a tamper-evident audit log.
Change the base URL. Your code, keys and infrastructure stay yours.
The measured difference
Less spend. More control.
Measured at a live customer. Not a projection. A 7-person team, over approximately two months.
−74%less on the invoice
30,297 requests
Approximately 3.65B tokens
One measured deployment—not a guaranteed saving. Results vary by workload.
- 01
Repeat work isn't billed twice
A repeat request never reaches the provider. It's still logged.
- 02
Context stops snowballing
Long agent loops stop paying for the same tokens every turn.
- 03
Tools stay out of the prompt
Capabilities are carried by Torii, not re-sent to the model.
- 04
Premium models where they matter
Allow-lists keep the priciest models for the people who need them.
- 05
Runaway loops hit a ceiling
Hard caps per user, team or key stop a loop before it burns the budget.
And more
Many smaller optimisations run on every request, too.
About these numbers
The comparison uses the existing measured customer sample: 7 people, approximately two months, 30,297 requests and approximately 3.65 billion tokens. The 74% figure is rounded.
Agent-heavy, repetitive workloads tend to benefit more than short, one-off requests. These figures compare the provider-spend amounts shown above, not a complete estimate of deployment costs or Torii pricing.
Deploy & connect
Single host
One installer. Preflight, keys and TLS included.
1command
$ curl -fsSL https://get.torii.io/install.sh | shKubernetes
One Helm release. The gateway scales on its own.
1Helm release
$ helm install torii oci://registry.torii.io/toriiOpenAI SDKs
Speaks /v1/chat/completions. Point the SDK, keep the code.
1line changed
$ base_url="https://torii.acme.com/v1"Claude Code
Speaks /v1/messages. Sees Torii as the Anthropic API.
0code changes
$ export ANTHROPIC_BASE_URL=https://torii.acme.com
Your infrastructure. Your rules.
Torii runs where your data already lives. No provider keys in the box, and no traffic through ours.
Talk to usIdentity
OIDC · SAML · SCIM
Your directory decides who gets in, and who is out the day they leave.
Audit
Tamper-evident
A hash-chained log you can verify from the panel in one click.
Providers
Your own hardware
Live now, alongside Bedrock and Vertex AI. Azure OpenAI is next.
Questions, answered.
Short answers to what teams ask first. For anything else, write to hi@toriigate.ai.
What is Torii?
Torii is a self-hosted enterprise AI gateway. It sits between your teams and every AI model they use, and prices, budgets, access-checks and audits each request before it moves on.
How long does it take to connect?
Only the base URL changes. Claude Code uses ANTHROPIC_BASE_URL and a token; OpenAI SDKs use base_url and api_key. No code, library or request schema changes.
Which providers and models are supported?
AWS Bedrock, Google Cloud Vertex AI and models on your own hardware are live, and Azure OpenAI is next. That covers Claude, GPT, Gemini, Llama, Mistral, Cohere and Amazon Nova, plus speech, image and embedding models, all under the same rules.
Where do our prompts and provider keys go?
Nowhere. Torii runs on your infrastructure, so prompts and credentials stay inside your perimeter. Provider keys are stored encrypted in the panel, never in a plain-text file.
How do we know the cost numbers are right?
Token counts come from the provider's reported usage, and every request is verified on top. Input, output, cache and reasoning are priced separately, and a model with an unknown price is flagged, never written as $0.
Does the 74% saving apply to every team?
No. It was measured on a 7-person team over about two months (30,297 requests). Agent-heavy workloads with a lot of repetition save more; short one-off requests save less.
Does the audit trail satisfy regulators?
Torii keeps a tamper-evident, append-only ledger per account, in line with EU AI Act Article 12, that you can verify from the panel or the API. Emails, cards, IBANs and API keys are redacted before anything is stored.
Does it work with our SSO?
Yes. People sign in through your identity provider over OIDC or SAML 2.0, and SCIM 2.0 provisions and removes accounts automatically, including revoking keys the day someone leaves.
Can we run Torii on Kubernetes?
Yes. One Helm release brings up the gateway, panel, data layer and ingress, with secrets generated inside the cluster. A single-command installer covers plain hosts.
Early access
Bring your AI
under your control.
Torii is opening to a first group of teams. Leave your email and we'll write once, when your access is ready.
Email first, then a few team details. We’ll write once when your access is ready.
— The Torii team
From the blog
All postsProduct & architecture
Why We Built Torii: The Control Problem in Enterprise AI
AI use inside a company grows one team at a time. This is why we built a single control layer for access, cost, data flow and audit, including for the models a company runs itself.
8 min read
Read the post