Klad gives coding agents real brushes: every stroke can be edited by hand afterward

If you have tried building a game with an AI agent, you know where things fall apart. The agent writes code with gusto, but art is where the pain begins. There are two paths: a generative model, where every edit turns into a new roll of the dice, or vector graphics that the agent draws with code, and which look exactly like what they are. The author of Klad got stuck right there and solved the problem with a tool.

The idea

Klad is a drawing app written in Rust that an agent controls through the command line. Everything available to the agent is also available to a human in the regular interface. Inside are a brush engine, frame-by-frame animation with bones and IK, vectors, text layers, colors, layer effects and infinite undo. Undo works even after a file is reopened.

The key architectural decision is editable strokes. An image is stored not as a flat set of pixels but as a sequence of strokes that are rendered as paint. So the agent produces a first draft with real brushes, and you step in and change any stroke or layer. Don't like the position of one line on a character's head? You move it, rather than regenerating the whole image and hoping the rest stays the same.

How to try it

The barrier to entry is low. According to the author, it is enough to clone the repository, show it to Claude, ask it to build the project and pick up the drawing skill that lives in the repository. After that you can ask for any drawing. The author tested the combination with Claude (the post names the version as Fable 5.2) and considers the results already impressive, but he candidly notes that quality depends on the agent. A weak model will cope worse with the CLI.

Here is the platform picture. On Mac the app works, and the author uses it every day. On Linux it builds, but tablet pen pressure sensitivity is still unfinished. Windows builds in CI; the author hasn't tested it there himself, although there are reports that everything works. Next he promises pixel-art support, which, by his account, will appear in the repository this coming weekend.

License and context

Klad is not open source but source-available. It is free for individual developers and small companies: you can build and use it. Later the author plans to sell ready-made builds, as Aseprite does. The file format, the CLI reference and the agent skill are under MIT. That is an important nuance: the format is open, so you are not locked into a single app, and other tools can read it.

The author ties the project to a trend: at Adobe MAX, he says, agents editing Photoshop files were shown, with a closed beta planned for early 2027. Klad is built on the same basic idea and is already available. The Hacker News post gained about 80 points and 32 comments, so there is interest, but this is still a niche project rather than a mass-market product.

Limitations and what to do

It is worth looking at the risks soberly. The project is run by one person, the license is not fully free, and the quality of the drawings depends directly on the model. There are no independent comparisons with generative images, so you will have to judge the result on your own tasks.

The practical takeaway is broader than the app itself. If you are building a tool that agents are meant to use, repeat this approach. Expose all functions through a CLI, describe them in the documentation, and put a ready-made agent skill alongside. Store the result as editable operations rather than a finished artifact: then a human can step in at any point. For indie game developers, it makes sense to spend an evening checking whether your agent can handle the CLI without hints and how easy it is to fix its strokes. That is exactly what the author asks people to share in the comments.