What shipped in the Blender agent skills wave?

Three Blender agent skills born Sep 7, 2026 — holo-card-studio (1,202 stars), dream-loop (407), holo-card (192), 1,801 combined by Sep 9 — all on GitHub's daily trending boards. Two are independent implementations of the same Xiaohongshu tutorial.

What recipe do the Blender agent skills share?

Image generation dreams a target screenshot, the agent builds toward it over headless Blender (MCP or scripting, preferred to computer use), and a separate critic AI compares live output to the target and loops until satisfied — with optional re-dreaming of a better target.

Why do Blender agent skills matter beyond demos?

They ship editable project files (card.blend, Blender scenes, Three.js pages) rather than throwaway images, install via one command like npx skills add, and scale to manufacturable work — earthtojake/text-to-cad applies the same skill pattern to CAD, CAE and CAM at 14,814 stars.

Blender Agent Skills: 1,801 Stars in Two Days

The hottest repo format in OSS right now is not a library — it's a skill. Three Blender agent skills, all born on Sep 7, hit 1,801 combined GitHub stars by this morning's scan: holo-card-studio (1,202), dream-loop (407), holo-card (192). All three sit on today's daily trending boards. I pulled the numbers and timestamps from the GitHub API minutes ago — this is the wave as it stands.

Three repos in 16 hours, one tutorial

The timestamps tell it: holo-card-studio went up 00:41 UTC, dream-loop 14:19, holo-card 16:18 — three working skills inside 16 hours. The receipt that ties them: holo-card-studio and holo-card credit the same Xiaohongshu tutorial handle (@乌托邦的香蕉, also on Bilibili), and holo-card explicitly calls itself an "independent implementation." One Chinese craft tutorial became two OSS implementations and 1,394 stars in a day. dream-loop (by @anshuc) is the genericized sibling: a skill for "impressive visuals" in any game, app, or scene.

Dream, build, critique — over headless Blender

dream-loop states the loop plainly: (1) image generation "dreams" a target screenshot, (2) the agent builds toward it — Blender via MCP or scripting, "preferred to computer use," (3) a separate critic AI compares live output to target and loops until satisfied, (4) optionally re-dreams a better target. holo-card-studio runs one sentence through four generated layers (subject / background / lineart / text), a Blender scene with parallax and holographic glow, and ships a Three.js page plus the editable card.blend. Distribution is one command: npx skills add achimala/dream-loop. dream-loop is tested with GPT-6 Astra in Codex; both holo skills are Codex skills living in ~/.codex/skills.

From holographic cards to manufacturable parts

Same boards this morning: earthtojake/text-to-cad — 14,814 stars, "a library of agent skills for CAD, CAE and CAM" — proves the pattern scales from holographic cards to parts you can mill. The skill layer, not the model layer, is now the fastest-moving surface in OSS. Deliverables are editable project files, not slop PNGs, and the critic loop handles quality. When a craft tutorial becomes three working skills, 1,801 stars, and one shared recipe inside two days, prompt-to-artifact has stopped being a demo.

Three repos in 16 hours, one tutorial

Dream, build, critique — over headless Blender

From holographic cards to manufacturable parts

Related