Every word an AI writes comes with a probability. Low ones are where it was guessing. Pick a real answer below and drag the threshold, or ask your own question further down.
The samples are real Llama 3.1 8B Instruct responses from examples/logprobs/. A word is flagged when any of its pieces has a probability (p = elogprob) below the threshold, exactly as the CLI does it. Hover or tap a word for its probability and the alternatives.
Only APIs that return token probabilities work here. Pick one, paste your key, ask something, and the answer shows up above with the unsure words marked.
Your key stays in your browser. This page has no server. It sends your key only to the provider you picked, directly from your browser, in the request that asks your question.
It is not saved anywhere unless you tick "Remember", and then only in this browser.
| Provider | Browser calls |
|---|---|
| OpenAI | Yes. Answers cross-origin requests from this site. Use a model that returns logprobs, like gpt-4o-mini. |
| OpenRouter | Yes. Allows any site. Only some upstream providers return logprobs, so the request asks OpenRouter to route to one that supports every parameter. |
| Together AI | Yes. Allows any site. Asks with Together's own "logprobs": 3 form. |
| Local Ollama | Not by default. Ollama refuses requests from other websites. Start it with OLLAMA_ORIGINS=https://hallucination-highlighter.vercel.app and use Custom with http://localhost:11434/v1. Your browser may also ask before letting a site reach your own machine. Or use the CLI, which has no such limit. |
| Anthropic | No. Its API does not return token probabilities, so there is nothing to highlight. |
Browser access checked on 2026-09-30 with a CORS preflight from this site's address. Providers can change this.
The CLI does this too, with presets for the same providers plus vLLM and Ollama, and no browser limits:
llm-token-visualizer --live "Who painted The Night Watch?" --provider openrouter llm-token-visualizer --live "Who painted The Night Watch?" --provider ollama --model qwen2.5-coder:14b