huashu-mac-use open-sourced: give any agent near-GPT-6 computer-use skills
AlchainHust · x · 2026-09-06
The author open-sources huashu-mac-use, aiming to give any agent computer-use capability approaching GPT-6 Astra. His thesis: beyond coding, desktop and browser control is the critical 'last mile' for white-collar agent work.
He demos Claude Code driving Blender on two tasks:
- 100k-ton cruise ship: one prompt—generate multiple high-res ship photos via codex, then build a matching Blender model. The agent created a side-view photo, then chained right-front/left-rear/aerial views for consistency, measured dimensions from photos, and wrote a 400-line parametric modeling script (superellipse hull lofting, ten deck levels, recessed balcony walls). Task-to-delivery: 33 minutes, 8 iterations at 3-30s preview each. It even self-reported its gaps: missing streamlined bridge shell, stern railings, pool area, no wake spray.
- Chocolate donut: air bubbles, the fry line, half-dipped chocolate drips, and 300 randomly placed flat-lying sprinkles all reproduced.
More from coding & agent
- Same model, 63% vs 97% on ARC-AGI-3: 7 harness rules beat max reasoning — victor_explore · 2026-09-07
- Matt Shumer: I 4x'd my Pro/Max subs as agentic models remove the attention bottleneck — mattshumer_ · 2026-09-07
- GPT-6 Astra codes the geometry, Dreamina Seedance 2.5 renders the final footage — HeyAmit_ · 2026-09-07
- Grok Bot's topic-based agent form factor hailed as better than threads — brandon_galang · 2026-09-07
- Codex defaults to 3 subagents — one config.toml line raises concurrency to 64 — daniel_mac8 · 2026-09-07
- Teaching Qwen Next 3D sculpting in Blender by having GPT Astra coach it via MCP — LegacyRemaster · 2026-09-07