what is cavepainter fr?

Cave Painter is an experimental server that lets AI agents draw pictures using GIMP, the real desktop image editing program, instead of using generative or diffusion based AI image models. Agents call simple tools like create a canvas, draw an ellipse, add text, and export, and each of these calls goes to a GIMP process that stays running between commands. This means an agent can build up an image step by step with real edits, the same way a person clicking around in GIMP would, rather than regenerating a whole image from scratch each time. Under the hood, the project is organized so the core drawing logic does not depend on GIMP specifically, with GIMP plugged in as one possible backend. The available tools cover things like creating canvases and layers, drawing shapes such as ellipses, rectangles, and bezier curves, rendering text, applying filters like blur or glow, and exporting finished images as PNG files. A separate feature called Prometheus lets an agent watch a human working in GIMP in real time, reading dialog settings and canvas changes as the person paints, so the agent can learn techniques from watching rather than only acting on its own. The README also describes an experiment where three separate AI agents were each asked to draw a self-portrait using only these drawing tools, with no generative image model involved. A fourth AI then judged the resulting images without seeing any of the underlying code. The author notes that two of the three agents ended up with very similar background colors, which they suggest may point to shared bias coming from using the same underlying model provider. This project would interest developers experimenting with AI agents that control real creative software instead of prompting an image generator, or anyone curious about giving an AI agent direct, tool based control over an existing application like GIMP.