Providers, models and routing¶
Your apps call one gateway with one key. Behind it you decide which providers it can reach, which of their models are offered and under what names, and where each allowed call goes: to a cheaper model, to a trial of two, to the next model when a provider is down, or only to models that run in the region you chose. Apps don't change a line of code when you do.
Everything here is in Models & routing in the Control room's menu: Providers, Models and Routing. Region rules and compliance routes are made from Policies, because they serve a compliance duty.
Providers and models¶
| Task | What you gain |
|---|---|
| Add a provider | Connect OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Mistral, your own servers and more, and know where each one runs before you save |
| Find a provider's models and switch them on | Bring in what a provider offers, switched off until you decide |
| Add a model by hand | Offer an Azure deployment, a Vertex model or anything a provider can't list |
| Price a model | Every call priced: from the rate card, your own negotiated price, or free |
| Give a model an alias | Stable names such as fast-chat that outlive the model behind them |
| Switch a provider or model off | See whose keys lose what before anything changes |
| See what changed on a provider | Who changed which setting, and when |
| Say where a provider runs, or add it in another region | Region rules and prices that know where data is processed |
| Add your own area, or map a location | Your own regions, such as an office data centre, inside the areas rules use |
| Manage a provider's keys and connection | Rotate a provider key with no outage; Azure, Bedrock and Vertex credentials |
| Read provider health and circuits | See an outage the gateway is already routing around, and bring a provider back |
Model access¶
| Task | What you gain |
|---|---|
| Choose which models a team or agent may call | Each team and agent reaches only the models it needs, seen before it changes |
| See who can call a model | Answer "who can reach this model?" by team, agent and key |
Routing¶
| Task | What you gain |
|---|---|
| Send a model's calls to another model | Cut cost or switch vendor for a team without touching its code |
| Trial two models with an A/B split | Real calls, split by share, with calls, failures, speed and cost per arm |
| Build a fallback chain | Calls keep answering through a provider outage |
| Keep calls in a region Enterprise | A team, agent, application or environment that only reaches models in the EU (or any area you choose) |
| Add a compliance route Enterprise | A model always sent to its compliant deployment, with no exemptions |
| See which routing rule wins | For any key and model, the one rule that decides, and why the others don't |
| Stop all AI traffic | One button that stops every AI call on every gateway node, and resumes them |
How the pieces fit¶
flowchart LR
A[App calls a model] --> B{Access:<br/>may this key call it?}
B -- no --> R[Refused]
B -- yes --> C[Routing:<br/>compliance route, required rule,<br/>A/B split, default rule]
C --> D{Provider<br/>taking calls?}
D -- yes --> E[Model answers]
D -- no --> F[Fallback chain:<br/>next model in your order]
F --> E
Access always comes first: routing never lets a key reach a model it couldn't call, and never blocks one it could. Region rules and the organisation's list of models approved for use narrow what routing may choose.
Who sees what: organisation admins change providers and models; organisation viewers and Compliance officers Enterprise read them. Routing is open to organisation admins, viewers, group admins (for their own teams) and Compliance officers.