One endpoint to every model.
The evidence is produced on the way through.
MODELGATE is an OpenAI-compatible gateway to the major models — Claude, GPT, Gemini, Mistral, Llama and the open weights — where security, integrity, jurisdiction and EU AI Act evidence are applied to every call before the response returns. Inspection runs in-process with no network call and no database write in the request path: 0.28 ms at p95, measured, against a 100 ms ceiling. The model's own time is the only latency you notice.
Trust primitives BUILT · MEASURED
Each one answers a documented failure of an existing gateway. None is speculative.
Bring your own key
Your provider key travels in one header for one request and is never stored - there is no storage path in the code. The leading gateway holds your key for every provider; one breach there exposes every account. We cannot leak what we never hold.
Hard budget ceilings
Per-day, per-month, per-call euro ceilings, calls per minute, tokens per call. Checked before any byte leaves, from a pessimistic estimate; a ceiling of zero is a freeze. A looping agent hits a wall, not an invoice.
Honest failure codes
A provider refusing your key is our 502 upstream_auth_failed - never a 401 that sends you debugging the wrong credential. Every failure says client, customer, provider or MODELGATE, and whether to retry.
Graceful queueing
Past your concurrency slots a request is held for the window, then fails as queue_timeout with Retry-After. There is no 429 in this gateway. Provider throttling is retried inside the same window.
A published SLA
Availability computed from our own request ledger per tenant. Provider failures and your own policy refusals are excluded from both sides of the fraction. Zero traffic reads UNVERIFIED - never 100%.
The compliance layer BUILT · MEASURED
This is the part a regulated buyer cannot get from a routing product, and the part we are native to.
Jurisdiction-locked routing
Every provider carries its legal home and where it processes. You declare the blocs you permit; a route outside them is refused with the provider and jurisdiction named. Failover stays inside your set. DeepSeek and Qwen are in the catalogue so the lock is demonstrable, not theoretical.
Zero-retention with proof
Not a sentence on a pricing page: a retention statement per request - prompt and response hashes, bytes seen, bytes retained (zero) - signed post-quantum by a key in hardware custody. Verify it with the public key and your own texts; the content is not needed to verify.
EU AI Act Article 12 / 14 evidence
One hash-chained record per request, written after the response has returned so it costs the request nothing. Every refusal raises an Article 14 oversight event a person can acknowledge. The export carries denominators: records chained, events reached a human.
Sector overlays
Each overlay is a machine-checkable set of controls mapped to the article it answers, assessed against your live configuration. You get a number and a list of named gaps, never an adjective.
What is public today
Three surfaces need no key, so you can read the contract before you talk to us.
The gateway itself answers only a tenant key. Tenants are provisioned by us during this phase; the local open-weight integrity gate (hash, provenance, licence and a signed certificate before a downloaded model ever runs) is the next phase.
Pricing
Not published yet, on purpose. MODELGATE is priced as a subscription, never as a percentage of your inference spend, and the tiers are being set with the compliance officers who will use the Article 12 export — not by us in a room. If you produce AI Act Article 12 logging today by hand, we would like thirty minutes.
Route through the gate.
Bring your own keys. Set your ceilings. Lock your jurisdictions. Get the evidence on the way through.
Talk to usThe 114 doors