UGT: Use the GitHub Thing

You found it on GitHub. Now use it.

Give your AI agent any GitHub link and the job you want done. UGT gets the project running on your own machine, proves the job worked on your files, and hands you the result. Windows, Mac or Linux. No Docker to learn. No cloud bill.

How it works · Get started · What you can do · Why not just ask your agent? · The catalogue · Share your recipe · House rules

How it works

1. Say "Use the GitHub Thing", with a link and what you want. "Use the GitHub Thing: github.com/openai/whisper. Put subtitles on this interview." Your words, your files.

2. UGT gets it running. Any GitHub project, no recipe needed: your agent works it out with UGT and proves it on a known sample before it touches your files. If someone has already shared a recipe for that project and that job, UGT uses it instead, building it exactly as it was verified, or pulling it ready-built for the featured ones.

3. You get the result, not a terminal. The file lands in your UGT folder. Run it again in one line. Stop it, remove it, or delete its data whenever you like.

4. Share it, if you like. Once it works, UGT can submit your recipe to the catalogue, so the next person gets it fast.

Related MCP server: GitHub MCP Server

Get started

Four steps. No terminal.

1. Download UGT for Windows, Mac or Linux.

2. Open it. It sets up a private Linux engine, connects itself to the AI apps you already use (Claude, Codex, Cursor and others), and opens your new UGT folder. Windows asks your permission once, and may need a restart.

3. Put your file in the UGT folder.

4. Tell your AI to "Use the GitHub Thing", with the link and what you want:

Use the GitHub Thing: github.com/openai/whisper. Put subtitles on interview.mp4.

The result appears in your UGT folder. In Claude Code and other apps with slash commands, /ugt does the same.

Coming with the UGT app. Until it ships, use the developer preview below.

1. Claude Desktop: download the bundle for your computer from the latest release, such as ugt-mcp-windows-amd64.mcpb, and open it. Claude Desktop installs it in one click.

2. Claude Code, Codex, Cursor or VS Code: install ugt with one line, which sets it up and connects the AI apps you have. Claude Code and Codex get the ugt skill too. ugt update keeps it current.

   • Mac or Linux: curl -fsSL https://usegithubthing.com/install | sh

   • Windows, in PowerShell: irm https://usegithubthing.com/install.ps1 | iex

3. The engine: on Windows, an Apple silicon Mac and Linux on a PC, UGT sets up its own private Linux engine the first time it is needed. On an Intel Mac or Arm Linux, have Docker Desktop, OrbStack or Docker Engine running.

4. Ask, exactly as above.

What you can do with it

GitHub is the world's biggest shelf of free software, and most of it never reaches the people it could help, because it ships as code, not as an app. UGT turns any of it into something you can just ask for.

A few featured examples. You never need a recipe from the catalogue: without one, your agent works the project out the first time. A recipe just makes it faster, because someone already did.

Why not just ask your agent?

Your agent can read the code. It usually can't run it for you, and it starts from zero every time.

• Most of GitHub assumes Linux. On Windows and on a Mac your agent hits a wall. UGT brings its own private Linux engine, set up in one step, on Linux too: no Docker Desktop, no account, nothing to learn. On an Intel Mac or Arm Linux it drives the engine you already have.

• Proven once, quick for everyone after. The first agent to get a project working can spend an hour building and fixing: memory limits, wrong-architecture binaries, missing system libraries, a resolver that cannot find the internet. The recipe it leaves behind turns the next person's run into a download.

• Proof, not "it started". A container starting is not your job done. Every recipe carries a check that fails unless the output is real: subtitles that fit the recording, a video of the length you asked for, a mesh you can actually print.

• Your machine, your files. Projects run in a sandbox that cannot see your home folder, your keys or your other apps. A recipe that needs more, such as the internet, a folder or a device, says why, and UGT asks you before it runs. Cloud compute is offered only with its cost on the table.

The catalogue

Recipes are a shortcut, never a requirement. Each one is a project someone already got doing a job, shared so nobody has to work it out again. A recipe is a folder anyone can read, in recipes/: a compose.yaml that holds everything that runs, and a recipe.json that says what it does. One project can have many recipes, one per job.

Browse them at usegithubthing.com/recipes, or read their source in recipes/. Every one was verified on clean machines before it was admitted, and a new recipe works for everyone the moment it lands, with no update to UGT.

Verified means every service ran as written on clean amd64 and arm64 machines and passed both checks, and both checks rejected deliberately broken outputs. A featured recipe's image is the one that was verified; any other builds on your machine from the same pinned commit and bases.

recipes/openai/whisper/subtitles/
  compose.yaml    everything that runs: build, limits, mounts, network, and the four services
  recipe.json     what it is: project, commit, task, settings, the access it needs, licence
  Dockerfile      fetches the pinned source itself, so the build needs no clone
  check/          the check scripts

It is plain Compose, so there is no new language to learn, and every recipe runs anywhere Compose does. The recipe reference has the rules.

Share your recipe

Got a project doing something useful? Share it, so the next person gets it fast instead of working it out again.

Tell your AI "Use the GitHub Thing: share this recipe", or run ugt recipe submit. UGT checks it runs again from scratch, makes sure nothing personal goes with it (your files, paths and keys), and opens a pull request as you. It is then verified on clean machines with no secrets available, and admitted once it passes; a maintainer reviews any recipe that reaches beyond its sandbox first.

Coming with the UGT app. Until then, from the source:

go install github.com/auscaster/ugt/cmd/ugt@latest
ugt recipe new <owner>/<repo> <task>     # start one, or adopt the project's own Compose file
ugt recipe verify <folder>               # until it passes
ugt login                                # once: approve a code in your browser
ugt recipe submit <folder>               # opens the pull request as you

ugt recipe submit --dry-run shows exactly what would be sent first. CONTRIBUTING.md and the catalogue rules have the details.

Maintain a project? A "Run with UGT" badge is on the way: put it in your README, and one click opens your project in UGT.

House rules

• Your files stay yours. Inputs are copied in. Your home folder, keys and the container engine are never handed to a project unless its recipe says why and a maintainer reviewed it, or you agree.

• Run anything, hide nothing. A recipe can need the internet, a folder, a device or more. It declares each one, and UGT refuses anything it did not declare. A maintainer reviews it before the catalogue admits it; a recipe from anywhere else asks you before it reaches your computer.

• Proof or it didn't happen. No check, no success.

• You decide. Installs, admin prompts, big downloads, accounts, keys and cloud compute are always your call, asked plainly before they happen.

• Secrets never pass through your agent. UGT asks you for them itself, once, and they live in your system keychain.

• Credit travels with the code. Every recipe records its project's licence, and files adapted from upstream keep theirs.

License

Apache-2.0. Each packaged project keeps its own licence. The UGT name and logo are not licensed; see NOTICE. Security issues go to SECURITY.md.