Type a prompt. A reply streams in token by token, and every other token gets rewritten on the way: reversed, uppercased, scrambled, deleted. The bar under each token is how sure the model was about it. Runs entirely in your browser; no API key needed.
Your key stays in this browser. The page calls the provider directly from your browser; there is no server of ours in between. It is only saved if you tick "remember" (browser storage on this device). Any OpenAI-compatible endpoint works if it streams /chat/completions with logprobs and allows browser (CORS) requests; if it does not, the request is blocked and this page will tell you.
The mock model is a small script in this page, not a real model: it picks a canned reply and makes up plausible confidence numbers, so the bars are simulated. The token rules are the real ones: same tokenizer, same transforms, and the same rate rule as the Rust tool (token i is rewritten when floor((i+1)*rate) > floor(i*rate), so 0.5 means every other token). Hover a token for its numbers; click or tap a rewritten one to flip it back to the original. With your own key, the tokens and the confidence are the model's own.
every-other-token does this to a live OpenAI or Anthropic stream in your terminal or its own web UI, with perplexity, top alternatives, confidence gating, provider diffs, JSON and CSV export. Free and open source (MIT), written in Rust.
$ cargo install --locked every-other-token $ every-other-token "Why is the sky blue?" --provider mock