Playground

Try DirectInference without writing a client. The playground has four modes — chat, effort sweep, compare, and code — each showing the request type your call was classified as.

Chat

A normal chat surface against the DI endpoint. Type any model id to steer intent — it is treated exactly as it would be over the API and echoed back — and pick an effort level. Each turn shows a badge with the classified request type, read from the X-DI-Request-Type response header, and a cache badge when a turn benefited from caching.

Effort sweep

Send one prompt across several effort levels at once and compare the answers side-by-side. It is the fastest way to feel the cost-and-quality tradeoff on your own prompts before you commit to a level in code — see Effort.

Compare

Put DirectInference next to your existing endpoint. Paste a base URL and key for any OpenAI-compatible service and the playground runs the same prompt against both, side-by-side. Your comparison key is used only from your browser for that session and is never stored server-side.

Code

Turn the current prompt, model id, and effort into copy-paste code for the OpenAI, Anthropic, and Gemini clients plus curl. It reflects exactly what the playground is sending, so you can lift a working call straight into your app — model id unchanged.

Open it

Head to the Playground and use the mode switcher at the top. Everything you try there bills against your balance like any other request and shows up in your usage.