Writing about OneStar

The full story

A popular app collects thousands of reviews a month and the store boils them down to one number. The number moves, and it never says what moved it: the bug that came with the last update, the paywall that pushed out long time users, the feature a whole audience keeps asking for. The reasons are all in the text. Reading it well, at scale, is the hard part, and the tools that did it were priced for teams with a procurement step.

Jared Watson built OneStar after trying the two things everyone tries. Reading reviews by hand finds a pattern after forty of them and misses most of the rest. Pasting them into a chatbot returns a confident summary with numbers the model estimated, which change between runs and cannot be traced back to a quote. OneStar takes a third route. Jev, a model from TypeSafe built for typed judgments, answers twelve fixed questions of each review on its own: is there a complaint, which theme, how severe, does the reviewer say they are leaving, is there praise, what do they ask for. Each answer is one option from a fixed list with a probability, never free text. Plain code then counts those answers into shares, trends, audience splits and the impact of each release, so a figure is exact for the reviews read and can be rebuilt without asking the model again. On Max, an LLM then writes short passages about those counts, such as what to fix first and who feels it most, and code fills in every figure and quote it uses.

Two things follow from that design and both are unusual in the category. Every number in a report opens the reviews it was counted from, with the model's answer and confidence beside each one, so anyone can check the work before acting on it. And when the model is not confident, the review lands in an unsure bucket that stays on screen rather than being forced into a category. Each question is scored against labeled sets of 300 real reviews per kind of product before it ships, and the cutoff for when an answer counts as sure comes from those scores. The method page shows the result per question.

OneStar reads reviews for any app or game, so a developer can run it on a competitor as easily as on their own product, and Compare puts up to four side by side on the same themes. Steam is covered alongside the two mobile stores, with review bomb days flagged by their volume and complaints split by playtime. Developers who connect App Store Connect or Google Play Console get every country and the full history of their own app, alerts when a complaint spikes or a rating drops, a weekly digest, Slack, and an inbox for replying to reviews with the theme Jev found in each one in view. Reviewer names are never read, raw review text never touches disk, and only excerpts of up to 200 characters are stored.

Public reports on well known apps and games are free to read without an account, which is how most people first see it. A free account runs three reports a month with no card. Pro, at $19 a month, is priced for one person rather than a procurement team. The site has been public since September 23, 2026 and launched on Product Hunt on October 8, 2026.