Platform
Situra sits between your applications and the model providers. It speaks both the OpenAI and the Anthropic dialect, picks an allowed route for every request and records what happened — without keeping what was said.
An illustrative request from a project with ES residency: Azure’s edge in Spain, the gateway’s checks, route selection and metering. Step through it at your own pace.
1Only the base URL and the key change. The rest of your code stays the same.
from openai import OpenAI
client = OpenAI(
base_url="https://api.situra.ai/v1",
api_key="situ_live_…",
)
r = client.chat.completions.create(model="gpt-5-mini", …)The official OpenAI and Anthropic SDKs, the frameworks built on them and coding agents work by pointing the base URL at Situra. Errors come back in the caller’s dialect.
| Method | Path | Compatible with |
|---|---|---|
| POST | /v1/chat/completions | OpenAI Chat Completions, including streaming, tools and images |
| POST | /v1/embeddings | OpenAI Embeddings |
| GET | /v1/models | OpenAI List models — only the models the key’s project may use |
| POST | /v1/messages | Anthropic Messages, including streaming, tools and images |
| POST | /v1/messages/count_tokens | Anthropic token counting (estimated if the upstream lacks it) |
Authenticate with Authorization: Bearer situ_… or, Anthropic style, x-api-key: situ_….
Any catalogue model can be called in either format: the gateway translates requests and responses, streaming included.
The data plane never queries the database on the hot path: it works from a snapshot of keys, balances, limits and routes that the control plane pushes over NATS.
The tier never widens
A fallback can only be another route in the same tier or a stricter one. If none is left, the answer is a 503 (no allowed route) or a 502/504 (upstream failure after all fallbacks).
Every served response carries the tier and region of the route that served it (x-situra-residency, x-situra-route-region); every response carries x-situra-attempts. The optional x-situra-residency: es|eu request header can only tighten the project’s tier.
Format and checksum are validated, the key is looked up in the snapshot and its HMAC verified with the key’s pepper version.
The key has not expired and the organisation is active (or pending verification, with test keys only).
Requests per minute for the key; requests and tokens per minute for the organisation.
The model is in the intersection of the key’s and the project’s allowlists.
Positive balance (prepaid) or within the credit limit (invoiced), and no hard-stop budget exhausted.
Active routes the project’s tier allows, zero-retention if the project requires it, by priority and weight.
If the route fails in a retryable way (429, 5xx, timeout, connection) before the first byte is streamed, the next allowed route is tried.
Provider-reported tokens, cost at the route’s current price, latency and the route used.
By default we record metadata: who, which model, how many tokens, what it cost, how long it took and how it ended. Prompt and response logging is opt-in per project, with 0–365 days of retention, encrypted under your organisation’s data key.
request_idmodel and routeproject and keyresidency and regioninput and output tokenscache tokenscost in EURlatency and TTFTattemptsstatus and error typeBreakdowns by model, project, key, provider, region and residency; latency percentiles; CSV export; comparison with the previous period.
Keys look like situ_live_<id>_<secret>, with at least 256 bits of entropy in the secret. We store only an HMAC-SHA256 with a pepper held in Azure Key Vault.
situ_test_ test keys against a deterministic sandbox, at no cost.
There are good AI gateways and workspaces. Situra focuses on per-request residency, ENS and the Spanish public sector.
| Offering | What it offers (public description) | Where Situra differs |
|---|---|---|
| Situra this service | Managed AI gateway in Azure Spain Central (Madrid), compatible with OpenAI and Anthropic clients; ES ⊂ EU ⊂ Global residency enforced per project on every request. | Not certified yet: designed for ENS category Alta; ENS Media + ISO 27001 audit targeted for Q2–Q3 2027. Private beta in Q4 2026. |
| nexos.ai | Gateway, chat and agents, adoption observer; EU-hosted. | ENS as a certification target, Spain-hosted, Spanish-first sales and support. |
| OpenRouter | Broad model aggregator behind one API. | Per-project data residency, a certification roadmap, and enterprise governance (budgets, audit, four-eyes). |
| Open WebUI | Open-source UI with a gateway layer, operated by each organisation. | Managed, multi-tenant service with certification as a target. |
| Cohere North | Enterprise workspace and agents. | Model-neutral gateway, ENS as a target, focus on the Spanish public sector. |
nexos.ai publicly lists GDPR, SOC 2 Type II, ISO 27001 and ISO 42001; ENS was not listed when we reviewed it in September 2026. (source: nexos.ai)
Based on vendors’ public websites as reviewed in September 2026. If something has changed or is wrong, tell us and we will correct it.
The data plane is a Go service that embeds Bifrost’s provider adapters (Apache 2.0) at a pinned release; authentication, residency, budgets and metering are our own code.
Private beta · Q4 2026
Spanish public administrations and regulated organisations that want to use AI models with data residency under control. We work with each organisation on ENS categorisation, the DPA and the integration.