Live prompt tester
A scratchpad for prompts that talks to the real models. Pick OpenAI, Anthropic, or Google, paste your own API key, write a system and user prompt, set temperature and max tokens, and send. You get the model’s actual response and its reported token usage, straight from the provider — no sign-up, no proxy, no data leaving your browser except the direct call you authorise. It is the fastest way to see how a prompt behaves before you wire it into code.
When to use a live prompt tester vs. the provider’s playground
Provider playgrounds (ChatGPT, Claude.ai, Google AI Studio) are great for exploration, but they hide details that matter when you are building:
- Token counts — the playground shows responses but usually does not surface the exact prompt tokens and completion tokens separately. The live tester shows both so you can cost-model your use case accurately.
- System prompt isolation — the tester sends a separate system prompt as its own message, matching what your code will do, rather than mixing it with the conversation.
- Exact model selection — playgrounds sometimes use the latest model version automatically. The tester lets you specify the exact model ID so you test against what you will actually deploy.
- Temperature and max tokens — playground defaults vary and are not always visible. The tester forces you to set these explicitly so your tests reflect your production configuration.
Understanding the parameters
Temperature controls how much randomness the model uses when selecting the next token. At 0 the output is deterministic (the same prompt gives the same response). At 1 the output is more varied. For factual extraction or structured output, use 0 or a low value. For creative tasks, use 0.7–1.0. Most production applications use a value between 0 and 0.5.
Max tokens sets an upper limit on response length. Setting it too low truncates responses mid-sentence; setting it too high increases cost on long responses. For most tasks, a sensible maximum is 2–4× the expected output length.
How it works
You choose a provider and model, then supply your key in a password field. The tool builds the correct request for that provider — OpenAI’s chat completions, Anthropic’s messages endpoint (with the direct-browser-access header), or Google’s Gemini generateContent — and sends it directly from your browser with your temperature and max-token settings. The response text and token counts are parsed out and shown, with loading and error states handled so a bad key or a rate limit gives you a clear message rather than a silent failure. Nothing is stored: your key and prompt live only in the page until you refresh.
Tips and notes
- Test at your shipping settings. Run the temperature and max tokens you will actually use, not the defaults — behaviour changes a lot between them.
- Use a restricted key. Since the key lives in the browser tab, prefer a scoped or low-limit key over your main production secret.
- Watch the token usage. It is your cost signal; a prompt that balloons tokens on edge cases will balloon your bill too.
- CORS is the provider’s call. If a provider blocks browser origins you will see a network error — that is expected, not a bug in the tool.