FAQ

Do I send a model name in requests?

No. LatentKit is route-based. The API key's assigned route selects provider and model execution. Application code sends task payloads only.

Can I use OpenAI SDKs directly?

LatentKit exposes a unified /v1 API and official @latentkit/sdk / latentkit packages. OpenAI-compatible drop-in is supported for many chat workloads by changing base URL and authorization, but route-based fields still apply — do not send model selection when using LatentKit routing.

What is the difference between the SDK and the CLI?

The SDK goes inside your application code. It sends runtime AI requests such as chat, embeddings, image, speech, transcription, video, and queue jobs.

The CLI runs in your terminal. It can log in through the browser, show whoami, send smoke-test requests, tail logs, inspect traces, check usage, manage routes and keys, and run CI checks. See CLI.

Where do I get an API key?

Sign in at console.latentkit.com, open API Keys or AI Router, and copy the secret when the key is created.

Can I call LatentKit from the browser?

Store keys server-side. Browser clients should call your backend, which calls LatentKit with LATENTKIT_API_KEY.

Can AI agents use the CLI?

Yes. Use --json, pass --tenant explicitly for workspace-scoped commands, and keep secrets in environment variables. Do not paste raw API keys or provider secrets into prompts. See CLI for AI agents and automation.

What is response_profile?

An optional hint (fast, balanced, deep) for speed vs depth when the assigned route allows request overrides.

Is there a public API playground?

Not in this docs site. Use the logged-in console AI Router and Developers sections to run real workspace-scoped test requests.

How do I know which provider handled a request?

Application code should not choose providers directly. Check response metadata and console request logs to see which route attempt succeeded. Workspace members can change routing in the console without redeploying the app.

Where is the full endpoint list?

The REST API overview lists every public /v1/* endpoint with a link to its reference page: chat, completions, vision, embeddings, images, audio/STT, speech, video, and the async queue. All endpoints share the same auth and routing model.

Which models and providers are supported?

See Models & providers for the supported provider list and capability matrix. The Providers section in the console shows the live model list your workspace can route to.

What happens when a provider is down?

Published routes define ordered fallbacks. When an attempt fails with a retryable error, LatentKit tries the next eligible route entry automatically — your app still receives one response, and response.provider reports which attempt won. If every entry fails, you get a 503 NO_HEALTHY_PROVIDER error with diagnostic detail.

How is usage billed?

BYOK routes are charged by your provider account. Platform Access routes deduct workspace credits using the prices shown in LatentKit. Successful responses include cost_usd for reconciliation. Availability and plan limits for each modality are shown in the console.

How do vibe coding prompts work?

Open Build with AI (Vibe Coding), pick your AI tool and stack (Next.js, Laravel, FastAPI, etc.), and copy a safety-aware integration prompt. No live key injection — set secrets in your environment after copying.