Most people who find something useful on GitHub never get it running. The README assumes a Linux machine, a working Python, the right CUDA, three services and an afternoon. Your AI app can read every line of the code and still not run it for you, because nothing on your computer is ready for it, and it starts from zero every time you ask.
UGT closes that gap. Install it once and your AI app gets a recipe for the project, a Linux engine to run it in, and a check that tells it whether the result is real. Ask “use the geek thing to turn this recording into subtitles” and the subtitles land in your UGT folder.
A recipe is the work done once
A recipe is a small folder: a Compose file that builds the project exactly as it was proven, a record of what it takes and gives, and a check that fails unless the output is genuine. Subtitles that fit the recording; a mesh you can actually print; a page that answers. A container starting is not the job done.
The first agent to get a project working may spend an hour on it: memory limits, binaries built for the wrong processor, a system library nobody mentioned, a resolver that cannot find the internet. The recipe it leaves behind turns the next person’s run into a download. Where no recipe exists, your agent writes one, proves it against a known sample with ugt recipe verify, and only then uses it on your files.
The engine comes with it
Most of GitHub assumes Linux. UGT brings its own: an Alpine layer with Docker Engine, built once in CI and run the same way everywhere. Windows imports it into WSL 2. On an Apple silicon Mac it boots as UGT’s own machine on Apple’s Virtualization framework, in about two seconds; on Linux it runs under QEMU with KVM, as you, with nothing else to install. Your folders reach that machine only while a command needs them.
Every recipe in the catalogue was verified inside that same layer on clean amd64 and arm64 machines, so it runs on exactly the engine it was proven on.
Contributions admit themselves
The catalogue is a folder of the repository, and adding to it is a pull request. UGT’s catalogue service reads each one, verifies the recipe on GitHub’s machines in a workflow that holds no secrets, checks that its check rejects broken output, has a separate model review an outside contributor’s recipe, and merges it when everything passes. Nobody waits for a maintainer to wake up; a recipe that asks for more than the internet, such as a folder or a device of yours, waits for one.
Admission takes only the delta. A recipe’s change is read, verified and served on its own, and nothing else in the catalogue is touched or rebuilt.
Open it from a web page
A project’s page on usegeekthing.com can hand it straight to UGT with an Open in UGT link. UGT’s own page opens on your computer, shows what the recipe runs, what it gives you and what it can reach, and runs nothing until you press Run.
curl -fsSL https://usegeekthing.com/install | sh
On Windows, irm https://usegeekthing.com/install.ps1 | iex does the same. Setup connects the AI apps you already use; from then on, the asking is theirs.
What it will not do
A recipe from outside the catalogue that reaches past the internet runs only once you agree to what it asks for, in UGT’s own prompt. Catalogue recipes run without asking, because their review is the agreement. When a project cannot work on your computer, your agent says why and what the options are, rather than pretending.
UGT, Use the Geek Thing, is Apache-2.0, and the whole path from a pull request to a run on your computer is in the open. It is independent and not affiliated with GitHub.