Google has put streaming behind its lightweight gateway. In Streaming comes to API Gateway: real-time and AI-ready APIs, without changing your spec, product managers Mak Ahmad and Geir Sjurseth and engineer Jeremy Whitlock announce Public Preview support for Server-Sent Events, chunked and NDJSON response streaming, WebSockets, and gRPC server-side and bidirectional streaming. Streams can stay open for up to one hour. The pitch is that nothing about the contract moves: “The same document that describes your existing API describes your streaming API.” Opting in is one flag on gateway creation, and “that single flag is the entire difference.” Gateways also get a deterministic hostname built from the gateway ID, project number and region, so a URL can be committed to Terraform before the resource exists.
The claims are architectural rather than numerical, so there is little to discount, but the Preview caveats are the part to read twice. Streaming is enabled only at creation time, so an existing gateway cannot be converted. Custom domains through Serverless NEGs, console enablement and Terraform support are all absent at launch. Request latency for a stream now measures the life of the stream rather than time to first byte, which will reshape every latency graph a team has. And the argument that a split estate of governed request-response APIs and ungoverned streaming endpoints is how “organizations end up unable to answer basic questions about who is calling what” is true, and it is Google’s reason for routing Gemini, Claude and gpt-oss model traffic through the same door.
The catalog holds Google Cloud API Gateway as a single API, the Google Cloud API Gateway Projects API, which is the management plane the post’s gcloud command drives: the APIs, configs and gateways that this flag now creates in streaming mode. The agentic access profile maps 11 operations, 6 of them acting, which is the create, update and delete surface the announcement touches. Spec presence, agentic access, auth clarity, the rate limit signal and operation distinctiveness are lit on the Agent Readiness score.
The Kin Score is 56.2, strong band, carried by access clarity at 78.9 and discoverability at 66.1 against contract governance at 9.8. The Agent Readiness score is 19.6, agent-aware, and the unlit dimension that matters here is event surface described. The post’s whole argument is that streaming needs no new description, and the catalog agrees in the least flattering way: it has no description of any streaming, eventing or long-lived surface for this provider at all. A gateway that will now hold a WebSocket open for an hour in front of an agent, with its MCP server dimension also unlit, is making a promise about governance that its own contract does not yet say out loud.