Answers you can play with, not paragraphs you scroll past
CreativeGPT is a generative answer engine. It takes a question, searches the web when the answer depends on the world, and composes a small interactive page instead of a wall of text — then shows you the sources it used.
How an answer is made
- Your question is sent to Gemini together with a system prompt generated from CreativeGPT's own component library, plus a randomised composition brief for this specific answer.
- Gemini searches when it needs to using the Google Search tool. The queries it runs are shown under the answer, not hidden.
- The model writes OpenUI Lang — a compact language where every line declares one component — and the browser renders it as it streams, so the page assembles in front of you.
- Real sources are attached. Grounding links are collected, redirects resolved, and publisher domains listed beneath the interface.
Why two answers never look the same
A model asked to be creative still converges on the same layout. So every request receives a concrete brief: a theme, an opening block, the mix of components it may use, and a block count between two and six. There are seventeen recipes, from a spare chart-and-paragraph answer to a six-block page with a timeline, a table and a quiz.
The brief also constrains honesty: if a requested block cannot be truthful — no real quotation, no real photograph, no honest number — the model must substitute an equivalent block rather than invent data.
Twelve visual themes
Colour, typography, corner radius, texture and chart palette all come from one theme, chosen to suit the subject rather than the default.
- Midnightscience and technology
- Sandhistory, travel and craft
- Newsprintpolitics, opinion and archives
- Blueprintengineering, space and maps
- Risographart, music and culture
- Terminalcode, security and data
- Botanicalnature, food and health
- Solarenergy, money and sport
- Candyfun, celebrity and kids
- Emberconflict, drama and heat
- Glacierclimate, medicine and clean tech
- Monolegal, finance and serious topics
See them side by side in the theme gallery.
The component library
Answers are assembled from twenty-three components. The model chooses which to use; it cannot invent new ones.
Documents blocks
- Resume
- A complete, print-ready résumé rendered on-page with a real client-side PDF download.
- Letter
- A cover letter with the same layout, download and copy-as-text actions.
Editorial blocks
- Chart
- Animated bar, line, area and donut charts with hover readouts and a themed palette.
- Metric
- A headline number that counts up, with a delta and a one-line explanation.
- Table
- Compact rankings, specs and side-by-side facts.
- Timeline
- A dated rail for history, biography, launches and processes.
- MapPlace
- An interactive OpenStreetMap panel with a real marker and coordinates.
- Compare
- A two-sided decision table with a verdict line.
- Gallery
- A caption-and-lightbox image grid for real, public images.
- CodeBlock
- Syntax-tinted, copyable code samples with a language tag.
- Callout
- Insight, fun fact, warning or tip, kept visually distinct from prose.
- Steps
- A numbered how-to flow for recipes, workouts and procedures.
- Quote
- A large pull quote for a real, attributed line that deserves the spotlight.
Playable blocks
- Quiz
- Self-scoring multiple choice with per-question explanations and a finale.
- Flashcards
- A flip deck with shuffle for vocabulary, dates and formulas.
- SliderGuess
- Drag to estimate a number, then see how close you were.
- Vote
- A two-to-four way this-or-that that records your pick.
- Speak
- Pronunciation playback in any language using the browser speech engine.
Documents you can download
Ask for a résumé, CV or cover letter and CreativeGPT composes a real A4 document inside the answer — then renders it to PDF in your browser when you press download. Nothing is uploaded, nothing is stored: the document exists only in your tab and in your browser's local history. There is also an open-in-new-tab option if you would rather print from the browser's PDF viewer, and a copy-as-text action for pasting into an application form.
Documents are built with placeholders for anything you have not told it — [Company], [Month Year], [your.email@example.com] — rather than invented employers or dates. Replace those before sending.
Grounding and sources
When CreativeGPT says something current, it should be able to show where that came from. Grounding with Google Search gives the model live results; the app then surfaces the search queries and the publisher domains, and resolves Google's redirect links so a citation points at the real page. If sources disagree, the answer is expected to say so rather than average them into false confidence.
Privacy and limits
Conversations stay in your browser, in local storage, and you can erase them from the History panel. The only cookie is an opaque, hashed identifier used to count your questions. The default allowance is fifty questions per browser per day with a wider network ceiling; counters reset at midnight UTC and a failed request is refunded.
Attached files are read in memory, sent to Gemini with your question and then discarded — there is no upload bucket and no stored copy. Images and PDFs go to the model natively; DOCX, XLSX, CSV, Markdown, TXT and JSON are converted to text first. The limits are deliberately small: three files, 3 MB each, 30 PDF pages, about 2,000 spreadsheet rows.
Frequently asked questions
What is CreativeGPT?
CreativeGPT is a generative answer engine. Instead of replying with a wall of text, it composes an interactive page: metrics, charts, maps, tables, timelines, quizzes, flashcards and pronunciation buttons, arranged for the specific question you asked.
How is it different from a normal AI chatbot?
A chatbot returns prose. CreativeGPT returns an interface built from a component library using OpenUI Lang, a compact language that streams as structured components. You can hover the chart, pan the map, flip the cards and play the quiz without leaving the answer.
Where do the facts come from?
Answers are grounded with the Google Search tool in the Gemini API. When the model searches, the app shows the queries it ran and links the real source pages below the answer, with redirect links resolved to the actual publisher.
Which model does it use?
By default CreativeGPT uses Gemini 3.5 Flash-Lite, the cheapest current Gemini model that supports Google Search grounding. You can switch to a larger model such as Gemini 3.8 Flash by setting the GEMINI_MODEL environment variable.
Why does every answer look different?
Each request receives a randomised composition brief: a theme, an opening block, a required mix of components and a block count between two and six. That is what stops every answer collapsing into the same three-metric-plus-chart layout.
How many questions can I ask?
Fifty questions per browser per day, with a wider network-wide ceiling to prevent abuse. The counter resets at midnight UTC and a failed request is refunded. A built-in demo answer does not count against the limit.
Can I upload files?
Yes — up to three per question: images, PDF, DOCX, XLSX, CSV, Markdown, TXT or JSON, 3 MB each and 4 MB in total. PDFs are limited to 30 pages and spreadsheets to about 2,000 rows so they fit comfortably in the model's context. Files are read in memory and sent to Gemini with your question; nothing is written to a database, and nothing is kept after the answer.
Is my chat history stored on a server?
No. Conversations stay in your browser's local storage and can be deleted from the History panel at any time. The only cookie is an opaque identifier used to count your daily questions.
Can the answers be wrong?
Yes. Generative interfaces can contain mistakes, and a confidently rendered chart is still only as good as the data behind it. Every answer exposes its generated source, links its sources, and carries a reminder to check anything that matters.
Does it work on a phone?
Yes. Components reflow to a single column, the composer docks to the bottom of the screen, and the interactive blocks use touch-friendly targets. There is also a light and a dark theme for the surrounding site.
Is CreativeGPT free?
The app itself is free to use within the daily allowance. Running your own copy requires a Gemini API key from Google AI Studio, and you pay Google directly for model usage.
Who makes it
CreativeGPT is built by Shah as part of GenesisAGI. The interface is composed with OpenUI and answered with the Gemini API.