The tools here work internally. They’re here for you to try them out.
Everything runs on local models, on hardware in one house.
Tools built for the workshop's own work, published as they are.
Every tool on the Labs page earns its place on internal work first: a compliance pass before a page ships, a heuristics read on a layout, and a bio drafted the week before a conference. They kept getting used, so they are here. Anyone can run one right now.
Each tool is in working order, with its wait and its limits on the page.
When the hardware is busy it says so. When a model has to wake up it says that, and a demo is labeled a demo. The seconds quoted on each page are measured on the machine that answers, so what a tool costs is what it costs.
Everything runs on local, open-weight models.
One graphics card answers at a time, so a tool will sometimes say it is busy and tell you how long. Nothing leaves the building.
The constraint sets what a tool is allowed to cost. If a tool takes longer than it is worth, it does not get built.
The same hardware runs the properties on the Launches page.
Among them: a daily two-voice show, a weekly editorial letter, a plain-language reader for federal executive orders, and a five-minute lesson series. There is a radio station with its own music map, a spirits catalog read by a vision model, a comics collection, and a delivery tracker. Each one started as something on this bench.
The full list is on Launches, with what each one does and where it runs.
Nothing here is for sale.
There is no account, no subscription, no trial, no waiting list, and no support commitment. Tools appear when they work, stay while they are useful, and get pulled when they are not. The Labs page has had things removed from it for that reason.
The Labs page is a selection. Other tools run here that only make sense against our own data.