Case study · Ripeat
Side projectProject period2026 – Present
Human-approved recipe generation
Built a pipeline that resolves a tool across package managers and GitHub, drafts a recipe under fixed rules and activates it only after human approval; agents can request a draft but never approve one.
01
Why
No catalog can cover every tool, hand-written recipes do not scale, and letting a model emit install commands is unsafe.
02
Decision
Ripeat resolves a name across Homebrew, apt, dnf, pacman, winget (including the Microsoft Store), Chocolatey, Scoop, Snap, Flatpak and GitHub, then drafts an ordinary recipe from the source it finds under fixed rules and presents it for review. It activates only after the person approves. When an MCP client requests a draft, ripeat prepares it and the person approves it in a terminal; the model never authors a recipe or supplies a source or checksum. Downloads require a published SHA-256, and vendor installers are verified by checksum and signature.
03
What I built
Built a pipeline that resolves a tool across package managers and GitHub, drafts a recipe under fixed rules and activates it only after human approval. Agents can request a draft but never approve one.
04
Trade-offs
- A draft covers only the machine it was produced on, since that is all ripeat can verify; other systems are added by someone on them.
05
Outcome
An unknown tool becomes a reviewed, reusable recipe in one step, and shared packs and bundles pass through the same strict checks.
Technologies
- Package Managers
- Supply Chain Security
- Human-in-the-Loop
Go