JevGPT

JevGPT is a chat assistant that writes every reply one word at a time. Instead of a model that generates whole sentences, JevGPT is built on TypeSafe's Jev System One model, described on Product Hunt as a model that can't write. In JevGPT, that model is put to work in a chat interface that opens with the question "What can I help with?", and each word of each answer is picked by Jev from a vocabulary of 1,772 words, one probability distribution at a time. The app is built around people who already hold a TypeSafe API key and want to run a conversation on their own Jev credits, at roughly a cent per reply. For years, the framing goes, large language models built to write text have been used to make choices. JevGPT returns the favor: a chat app where TypeSafe's Jev, a model built to make choices, writes every reply one multiple-choice word at a time. The Product Hunt tagline states the premise bluntly — "A chatbot built on a model that can't write" — and the Product Hunt description answers its own question, "Does it work? Sort of." That framing makes the project less a polished productivity tool than a visible experiment in what happens when a decision-making model is asked to produce written text. The central mechanic is word-by-word generation. As the site states, "Every word is picked by Jev from a 1,772-word vocabulary, one probability distribution at a time." Rather than emitting a full sentence in a single pass, the app resolves each next word as a choice over that fixed vocabulary, then moves on to the next word. The vocabulary is small and fixed, and every reply is assembled sequentially from it. Watching this happen makes the generation process unusually visible: a reply is not retrieved as a block of text but constructed word by word, where each word is a multiple-choice decision made by the model. Because both the vocabulary and the selection step are constrained, the shape of the output is determined by that choice process rather than by open-ended text generation. JevGPT does not ship with its own model access. To start chatting, you enter your TypeSafe API key, and the app runs on your own Jev credits; the site states that a reply costs about a cent. After entering the key, you press Save and the session is ready to use. Keys come from TypeSafe's console — the site links to console.typesafe.ai, specifically the settings page for keys — for anyone who needs to create one. This bring-your-own-key arrangement means usage is metered against the individual user's own credits, and the cost of a conversation is expressed per reply rather than through a subscription described on the site. The site is also explicit about how the key is handled. It is "kept in an httpOnly cookie in this browser and sent only to this app's server." That statement covers both storage and transmission: the key persists in a cookie that page scripts cannot read, because httpOnly cookies are not exposed to JavaScript, and it is sent only to the server that powers this app. For anyone who hesitates to paste an API key into a web app, this stated handling is the detail that matters — the credential stays in the browser's cookie store and travels only to the app's own backend. If you do not have a key yet, the site offers two alternatives: watch the demo video, or read how JevGPT works on GitHub. Overall, JevGPT is a conversational layer over TypeSafe's Jev. The user supplies the credentials, the app passes the conversation to Jev through TypeSafe's API, and the model returns its selection for each word, a process the app repeats until a reply is complete. Because each word is drawn from the same 1,772-word vocabulary via a probability distribution, the output is shaped by that constrained choice process. The project is open source, with the GitHub repository linked directly from the site for readers who want to study how it works rather than watch it produce a reply. The site pairs that repository link with the demo video as the two routes to understanding the project without running it yourself. The stated benefits follow from those mechanics. You keep control of your own usage, since the app runs on your own Jev credits and a reply costs about a cent. You can inspect the project, because it is open source and the repository is presented as the place to read how JevGPT works. You can also approach it with honest expectations, since the Product Hunt framing answers whether the approach works with "Sort of" rather than a promise of polished prose. And because the vocabulary is only 1,772 words and every word is a discrete choice, the generated replies are constrained in a way that is visible to the person reading them, one word at a time. Concrete uses described in the content are straightforward. The primary one is chatting: the app opens with "What can I help with?", you enter your TypeSafe API key, and you start a conversation whose replies are written one multiple-choice word at a time. A second is evaluating the model: users can watch whether a model built to make choices can actually carry a written reply and form their own judgment. A third is developer review — the GitHub repository is offered so that people can read how JevGPT works, and the project is listed as open source. A fourth is passive exploration: the demo video lets someone see JevGPT operate without needing a key first, since the site suggests it to readers who do not have one yet. JevGPT is a browser-based web app, and its audience is narrowest at the point of access: you need a TypeSafe API key and Jev credits before you can chat. That points toward people already working with TypeSafe's console and looking for a hands-on way to see Jev in a conversational role. Beyond that, the project is open source and tagged with topics including Open Source, Writing, Artificial Intelligence, and GitHub, which speaks to developers and technically curious readers who want to examine the code. Pricing is usage-based by nature: JevGPT runs on your own Jev credits, and the site puts the cost of a reply at about a cent. No subscription tiers or plan details are mentioned in the content. The takeaway is that JevGPT is a deliberately unusual chat app: it asks a model built to make choices to write, and it writes by choosing. Every reply is assembled one word at a time from a 1,772-word vocabulary, one probability distribution at a time, on the user's own TypeSafe credits at roughly a cent per reply. It is open source, it documents how your API key is stored and sent, and it offers a demo video for those who do not have a key yet. Whether the writing is good is answered with a shrug — "Sort of" — and that honesty is part of the project's character.