An AI gateway in front of the providers can meter, route and log, but only by what each request tells it. Until now a request carried the provider's credentials and nothing about who in the platform caused it. This release gives every model call one origin, an identity, and limits the platform sets rather than the caller.
One place every request starts
- Headers pinned first A loopback test server records the method, path and headers of every call to each provider: Anthropic, Cohere, Ollama, OpenAI, Vertex AI and watsonx. The recording was green before the change and after it.
- One entry point Thirty-four request sites across the six providers now start from a single request builder. Each provider keeps its own authentication; a header every model call must carry is one line.
The caller, named to the gateway
- Three headers
X-Airtool-User,X-Airtool-DbandX-Airtool-Toolcarry the principal, the database and the agent or feature behind the call, taken from the call's bound context. - Only to the gateway They are sent only when the request goes to a base URL configured as the gateway. A value that is not plain ASCII travels as an RFC 2047 encoded word.
Models the platform chose
- Held to the catalogue A script that builds a model from a provider can name only a model that provider catalogues for that kind, and is refused at the point it names it, with the catalogued models listed. Before, an uncatalogued model reached the provider and failed at the first call.
- Locked where resolved A builder that stands for a resolved configuration, the user's default or the global model, refuses any model name at all.
- An output ceiling per model A chat model's configuration can set its maximum output tokens, which bounds tool-call arguments too. Anthropic's ceiling had been fixed at 8,192; a thinking budget now raises it to at least twice the budget instead of overwriting a larger setting.
Spend, routing and audit outside the platform can now be attributed with the same identity the platform itself records.