In short
AI generates utility apps, but their delivery and security remain a challenge. Krate bundles the code and permissions into a single cross-platform file, requiring the user to explicitly grant access before running the app.
AI is already capable of generating small utilities tailored to a specific task for an individual or a team. But delivering such code is a weak point: browsers don’t always handle local tasks well, and traditional desktop apps are tied to the operating system and often receive more permissions than necessary.
The Krate project solves this problem using the .krate format—a single file that contains both the application itself and a declaration of the access it requests. The file can be opened on Mac, Windows, and Linux, but before launching, Krate shows the user a checklist of what the application wants to access. The user approves only what they’re comfortable with.
This is an interesting shift in the distribution model. Instead of blindly trusting a downloaded binary or an installation via a package manager, permissions become part of the format. If AI agents start generating and sharing such files with each other, an explicit permission request will be the only barrier separating automation from uncontrolled access to the system.
The project is written in Rust (judging by Cargo.toml and rust-toolchain.toml), has 533 commits, and is licensed under MIT/Apache 2.0. However, its status is pre-alpha, and the macOS app has not yet been signed. In other words, the idea is strong, but it’s still a long way from production.
The main question isn’t whether Krate itself will take off, but whether explicit permission requests will become the de facto standard for any AI-generated applications. Without that, running code from an agent on your own machine is too big a risk.