Coroutines first: notes on nRouter's Kotlin SDK NRouter released a coroutines-first Kotlin SDK (ai.nrouter:nrouter-sdk-kotlin:3.1.2) for its API gateway, using suspend functions for calls and a cold Flow for streaming responses. The client reads NROUTER_API_KEY from the environment, dispatches on Dispatchers.IO, exposes gateway routing preferences like nrouter_optimize and nrouter_cost_band in the request payload, and surfaces x-nr-* header metadata as typed fields. Gateway refusals map to typed NRouterError subclasses with an isRetryable flag that is true for RateLimit, Service and Transport errors. A Kotlin client for an API gateway can be a thin wrapper around blocking HTTP calls, with a coroutine adapter bolted on. The nRouter Kotlin guide goes the other way: the SDK is coroutines-first from the start — suspend functions for calls, a cold Flow for streaming, and typed error subclasses you can branch on. Full guide: https://nrouter.ai/docs/sdks/kotlin https://nrouter.ai/docs/sdks/kotlin The guide publishes the artifact to your local Maven cache ./gradlew clean check publishToMavenLocal , then it's a normal Gradle dependency: dependencies { implementation "ai.nrouter:nrouter-sdk-kotlin:3.1.2" } NRouter reads NROUTER API KEY from the environment, targets https://api.nrouter.ai/v1 , and dispatches on Dispatchers.IO , so calls don't block your caller's context. For Android apps, the guide points to the separate nrouter-sdk-android artifact. suspend fun main { val client = NRouter val result = client.chatCompletions JSONObject .put "model", "nrouter/auto" .put "messages", listOf mapOf "role" to "user", "content" to "Hello " .put "nrouter optimize", "balanced" val content = result.body.getJSONArray "choices" .getJSONObject 0 .getJSONObject "message" .getString "content" println content println "Request ID: ${result.meta.requestId}" } Two things worth noticing. The routing preference rides in the payload: "nrouter/auto" plus "nrouter optimize" quality , balanced or cost , with "nrouter cost band" low , mid , high available under cost . And the response carries .meta extracted from the gateway's x-nr- headers, so request ID and cost metadata arrive as typed fields instead of header strings you parse yourself. Streaming is a cold Flow — nothing moves until you collect: client.messagesStream JSONObject .put "model", "nrouter/auto" .put "messages", listOf mapOf "role" to "user", "content" to "Hello " .collect { chunk - print chunk.delta } Cold matters here: the request starts on collection and stops when the collector goes away, which is exactly the lifecycle behaviour you want in a view model or a Ktor handler. Gateway refusals map to subclasses of NRouterError : Request 400, including cost band unsatisfiable , GuardrailBlocked 400 , Authentication 401 , Credit 402 , RateLimit 429 and Service 503 . The detail that makes this practical is isRetryable — true for RateLimit , Service and Transport — so a retry policy is one when branch instead of a table of status codes maintained by hand. That's the whole shape: a JVM client that treats coroutines as the default rather than an add-on, and lets the gateway carry routing, budgets and guardrails behind one key. The guide has the full walkthrough: https://nrouter.ai/docs/sdks/kotlin https://nrouter.ai/docs/sdks/kotlin