See your AI spend, cap it and cut it.
A provider-independent enterprise AI gateway. Accurate cost, per-user and per-group budgets, a tamper-evident audit trail, and 74% savings measured at a live customer.
Runs on your own infrastructure. Zero code changes: one line, the base URL.
Models reachable through one API
Providers
- AWS BedrockLive
- Google Cloud Vertex AISoon
- Microsoft Azure OpenAISoon
- Your own hardwareSoon
The core stays the same; providers are adapters that plug in and out.
Teams connect straight to the model.
Fast, but uncontrolled.
Claude Code, the OpenAI SDK and Cursor made teams faster. Four questions grew just as fast.
Why is the invoice this high?
Who spent what, on which model, for how many tokens? Direct access gives you no visibility. And several popular gateways get the arithmetic structurally wrong, billing usage served from cache at full price.
Who can reach what?
Everyone can call every model. Restricting the expensive ones, setting per-person or per-team limits and carving out exceptions stops being manageable by hand.
Where did the data go, and who did what?
There is no audit trail. Regulations such as the EU AI Act require tamper-evident records, and a log file assembled after the fact does not qualify.
Provider keys are everywhere.
Credentials for every cloud you use end up scattered across .env files and repositories. Unapproved tools and MCP servers turn into spend nobody owns.
Everything you govern, in one place
Rules are not written in config files that drift. You set them in the panel, and they are applied on the very next request.
This month
1–31 AugSpend
$2.184,60
12,4%
Tokens
418,2M
8,1%
Requests
9.412
5,3%
Daily spend
Budget alert
Design team reached 80% of its monthly ceiling
Spend, tokens and requests at a glance
Live totals broken down by model, provider, connection and user, defaulting to the current month rather than a confusing all-time pile.
We don't report cost, we reduce it.
Most gateways only report cost, and most of them report it wrong. The numbers below are not estimates: they come from the logs of a team running Torii, reviewed with their permission.
Same workload, two scenarios
7-person team · ~2 months · 30,297 requests · ~3.65B tokens
Savings
0%
$0.00 less on the invoice. Same team, same models, same output. The only difference is that the requests went through Torii.
You never pay twice for the same work
When a request that has been seen before comes round again, it never reaches the provider. The usage is still recorded, it just isn't billed; there is no hidden zero.
Context does not grow for no reason
In long conversations and agent loops the same content is billed again on every turn. Torii cuts that repetition without changing what the model returns.
Tool use doesn't inflate the bill
On workloads that lean on tools and capabilities, most of the cost comes from what gets carried along rather than the work itself. Torii keeps that weight outside the model.
One gateway, six jobs
Torii was not built for chat, it was built to govern how an organisation uses AI. Every rule you set in the panel is applied on every request.
Cost accounting
Every request is costed separately for input, output, cache and reasoning, then written to the ledger. Accounting runs outside the request path, so responses are never delayed by it.
- input
- output
- cache
- reasoning
Access control
Decide centrally who can use which model, on what budget, with which capabilities. "Sonnet for 50 people, Opus for 5 seniors" is set up in one place.
Chargeback and showback
Assign cost centres to users, pull spend reports per department and export to CSV. Team-level usage is aggregated server-side.
Capabilities: MCP and Skills
An admin curates the account catalogue; users enable an item or request approval. Per-user and per-group allow-lists control which model can reach which tool.
Audit and compliance
A tamper-evident audit log with integrity verification from the panel. PII redaction, data-residency control and tenant isolation at the database level.
Resilience
Automatic failover across model groups, staged retries and load distribution. If one region stops responding the request is routed to another automatically.
- model group
- automatic failover
- region routing
Zero code changes.
One line, the base URL.
Torii speaks both the OpenAI (/v1/chat/completions) and the Anthropic (/v1/messages) API surface. Claude Code sees Torii as if it were the first-party Anthropic API.
export ANTHROPIC_BASE_URL="https://gateway.yourcompany.com"export ANTHROPIC_AUTH_TOKEN="sk_live_..." claudeThe highlighted lines are the only thing that changes. The rest of your code runs untouched.
One command. 47 seconds. Your own server.
Self-hosted usually means a week of setup. Torii is a single installer: it runs the preflight checks, generates the keys on your machine and brings the whole stack up behind TLS.
$ curl -fsSL https://get.torii.io/install.sh | sh -s -- --key tk_setup_9f4c2ab7e1
Representative output. The setup key is single-use and is destroyed at the end of the run.
Nothing to wire up by hand
Gateway, panel, the data layer and a TLS-terminating reverse proxy come up together. Schema updates run themselves; there is no configuration file to hand-edit before the first request.
Keys are born on your machine
The master key and every database and cache password are generated locally and written to a chmod 600 file. None of them is transmitted anywhere, and the single-use setup key destroys itself once the run finishes.
No cloud credentials on the box
The installer never asks for a provider key. You add connections later from the panel, where they are stored encrypted rather than sitting in a plain-text file.
What's different?
Where Torii separates most clearly: it does not merely display cost, it reduces it architecturally. And accounting and budget correctness are a design decision, not a feature bolted on later.
Setup (Claude Code, SDKs)
- Direct provider
- Partial:Configured tool by tool
- General-purpose gateway
- Partial:Needs a bridge or compatibility layer
- Torii
- Yes:Native Anthropic and OpenAI compatibility
Cost arithmetic
- Direct provider
- No:After the invoice, allocated by hand
- General-purpose gateway
- No:Rounding error accumulates over time
- Torii
- Yes:Exact to the cent, no drift
Record integrity
- Direct provider
- Partial:Depends on the provider's invoice
- General-purpose gateway
- No:Known cases of usage line items silently failing to be recorded
- Torii
- Yes:Every request is verified; an unknown price is flagged, never hidden
Budget enforcement
- Direct provider
- No:None
- General-purpose gateway
- No:Reports of ceilings being passed through, counters inflating, and team spend charged to individuals
- Torii
- Yes:Two-phase budget (pre-flight and commit); fail-closed when resolution errors
Actively reducing cost
- Direct provider
- No:None
- General-purpose gateway
- No:Reports only
- Torii
- Yes:Cuts repetition, context bloat and tool overhead
Audit trail
- Direct provider
- No:None
- General-purpose gateway
- Partial:Records that can be edited after the fact
- Torii
- Yes:Hash-chained; tampering is provable
MCP and Skills management
- Direct provider
- No:None
- General-purpose gateway
- Partial:Tool schemas reach the model; reports of reconnecting per call and leaking slots
- Torii
- Yes:Managed catalogue, allow-lists, overhead control
Gateway overhead
- Direct provider
- Yes:No layer in between
- General-purpose gateway
- No:Measurable added latency; on the maintainers' own agenda
- Torii
- Yes:Accounting sits outside the request path; responses aren't delayed
Behaviour under load
- Direct provider
- Partial:The provider's SLA
- General-purpose gateway
- No:Reports of memory exhaustion, connection pool starvation and leaks
- Torii
- Yes:Predictable resource use; a small, pinned dependency set
Dependency surface
- Direct provider
- Yes:The provider SDK only
- General-purpose gateway
- No:A wide package tree; a published package was compromised once
- Torii
- Yes:A minimal dependency surface
Tenant isolation
- Direct provider
- Partial:Not applicable
- General-purpose gateway
- Partial:Application level
- Torii
- Yes:Isolation enforced at the database level
Key security
- Direct provider
- No:Scattered .env files
- General-purpose gateway
- Partial:Varies by deployment
- Torii
- Yes:Encrypted and centralised
The middle column is not hypothetical: these are classes of problem documented in the public issue trackers of widely used open-source gateways. Some of them will be closed over time. The distinction drawn here is not about individual bugs, but about the architectural choices that let them happen.
Questions about the enterprise AI gateway
If you can't find what you're looking for, write to us and an engineer will answer.
Take control of how your organisation uses AI.
Think of Torii not as a line on your cost sheet, but as the lever that lowers your invoice.
- Runs on your infrastructure
- Zero code changes
- EU AI Act ready audit