Fetchwright

About

Fetchwright is the fetch layer behind RoundupForge, a tool that turns public product and search pages into material for product roundups. It replaces a rented scraping API with software we own, measured against a benchmark that was frozen before any implementation was scored.

How Fetchwright was built is told in the lab notes. In short, the project runs as two public experiments. Experiment 1 builds Fetchwright itself. Two engineering agents build it independently — Fable, built on Anthropic’s Claude, and Astra, built on OpenAI’s models — against one frozen benchmark; the best parts of both get merged. Experiment 2 tests JEV, a classification model, as a relevance filter that keeps accessories and unrelated items out of product roundups.

Who runs it

Fetchwright is a Grimfaste experiment, directed by Thorsten Meyer in Iffeldorf, Germany, and published under the Thorsten Meyer AI studio. The name is an old trade word: a wright is someone who makes things — a wheelwright, a shipwright. A fetchwright makes the thing that fetches.

How we fetch

Fetchwright reads publicly available pages only, at a measured pace, and keeps what it has already read instead of fetching it again. It does not log in, does not collect personal data, and is not offered as a service to third parties.

Last updated 7 October 2026