BYOK playground

What this shows Not a general-purpose proxy: only /models and /chat/completions, https:443, no redirects, rate-limited. The key lives in this browser. DNS checks are best-effort against rebinding; platform egress is the backstop.

Your endpoint

The key is sent in the JSON body and is not stored on the Worker. Session storage is the default. One Turnstile pass mints a signed session cookie.

API key Base URL Model id

Remember on this device

Complete Turnstile once for a signed session. Compare uses action compare.

What you can try

Paste a key, an https base URL, and a model id, then stream a chat. Multi-turn sends the transcript (at most 15 turns, 4,000 bytes each — about 1,300 CJK or Devanagari characters).
List models from GET /models. The Worker returns only { data: [{ id }] }.
Send the same prompt to your model and to Workers AI, and read latency, time to first token, and tokens.
Probe streaming, tools, JSON schema, and vision. A pass is evidence-based (tool_calls, schema keys plus nonce, left-pixel colour); otherwise the check is fail, unsupported, or inconclusive.

What you could build

A product that accepts a customer’s OpenAI-compatible key without writing it to disk.
A migration bench that compares a provider against Workers AI on the same prompt.
An eval harness that records which capabilities an endpoint actually implements.

Streaming chat

SSE passthrough to /chat/completions. Chat is not stored on the Worker. Output is rendered as text, never HTML. A reply that ends without [DONE] is marked “stream ended early”.

System prompt

Temperature 0.4 Max tokens 256

Your endpoint

What you can try

What you could build

Streaming chat

Same prompt, two models

Yours

Workers AI

Capability checks