klad-app/klad

What it is

klad is a 2D painting and animation editor built for collaboration between a person and a coding agent. The person paints in the editor with a pen; the agent paints through the klad command line using the same brushes, layers and timeline, and the two see each other's strokes appear within about a second. There is no image model inside: the agent (Claude Code, Codex, or any tool that can run commands) decides what to paint, while klad supplies the drawing tools.

The central design choice is that everything stays editable data. A stroke keeps its path, pressure, brush and colour; a shape keeps its points; a fill keeps its region. Any stroke, human or agent-made, can later be selected, recoloured, bent, moved or deleted, and every change is one step in a shared undo history. The app also covers vector work (gradients, glows, text in bundled open fonts), animation with onion skins, clips, in-betweens, bones with IK, and game-oriented export to PNG, sprite sheets or a Godot SpriteFrames scene README.

It targets illustrators, animators and game developers who want agent assistance without giving up stroke-level control, and developers who want to script drawings through a CLI and documented file format.

Quick start

The project is built from source; there are no installers yet. You need Rust via rustup and just. On Linux also install pkg-config, libfreetype-dev, libfontconfig1-dev, libxkbcommon-dev and libwayland-dev.

git clone https://github.com/klad-app/klad && cd klad
just compile                  # about ten minutes the first time
target/release/klad open      # on Linux: target/linux/release/klad open

Then expose the CLI to your terminal and agent, and start the editor on a drawing:

target/release/klad install   # on Linux: target/linux/release/klad install
target/release/klad open samples/lantern/lantern.klad

To let an agent paint, start it in the folder where the drawing should live and give it a prompt that first makes it read the skills and then paint, for example: run klad skill and klad skill drawing, then create scene.klad and paint a night scene with a lantern and a small creature. Once the file exists, klad open scene.klad in another terminal shows the strokes land. You can pin notes with N and tell the agent to read them, or take over any stroke with S. Agent skills are MIT-licensed and printed by the klad skill commands Try it.

A first-start window offers guided steps and six bundled sample drawings; klad samples lists them and copies one out.

Constraints

• Licence: source-available under PolyForm Internal Use 1.0.0 with additional permissions, not open source. Personal use and building from source are free; a company with 500,000 EUR or more in revenue in its last tax year needs a paid licence per user, and redistribution of source or builds is not allowed. The file format, CLI docs, agent skills and klad-client crate are MIT Licence.
• Platforms: developed on macOS with pen pressure and tilt. It builds and passes tests on Linux and Windows, but on Linux only mouse input works for now (no tablet pressure) and nobody has drawn with it on Windows yet.
• Self-build: no signed installers exist yet; paid builds are planned for macOS, Windows and Linux.
• Feature gaps: some illustration basics are missing, including multi-shape selection, align/distribute, snapping, grid, and SVG import/export. Bones are a preview and their commands may change.
• Data: each drawing is a folder of plain files, needs no account, and only checks for updates once a day (which can be turned off), so drawings work well in git.

HotRepo's take

klad first surfaced on HotRepo's new-project list for the Rust language on 2026-11-11, gaining 13 stars in a window that ran for roughly one and a half hours and ending at 113 stars, with the show HN thread drawing 66 points and 28 comments that day. The discussion is more concrete than typical launch chatter: one commenter tested agent-driven drawing against a reference image and found the output clearly behind the classic image-to-video approach, with the model named as the bottleneck test report and author reply; another reported having an agent produce characters that look less obviously LLM-generated than usual and linked the results agent result. That tension — real, stroke-level control versus current model craft — is the thing to watch; the app's value depends on the agent's skill at drawing, not on the editor alone.