Simon Willison published a TIL (2026-09-05) showing GPT-6 Astra — OpenAI's new flagship model, released around Sept 3, 2026, with computer-use, code-execution, and MCP capabilities — running inside ChatGPT's macOS Codex mode and driving the locally installed Blender headless via `/Applications/Blender.app/Contents/MacOS/Blender --background --python scene.py`. The agent wrote the Python, Blender built and rendered a pelican-on-a-bicycle scene, and two follow-up prompts revised it (2m39s, 3m51s, 5m59s per pass); Willison published the scripts, .blend outputs, an exported Codex transcript, and a reusable 'Blender Local' Markdown skill on GitHub. Independent write-ups (Ciyo, kingy.ai) confirm the mechanism — Astra's model page lists text-only output, so all 3D artifacts are produced by Blender executing code the model wrote — and reference OpenAI's own architectural-visualization example. The case's broader corroboration (three Reddit practitioners running MCP-based variants for scene generation, multicolor 3D-print prep, and facial-mesh retopology) comes from the case's own evidence rather than these snippets, and the exact background-Python path has not been independently rerun.
Willison plus four independent practitioners keep newly arriving at a position Scott's canon already holds — the model writes Python that Blender executes between judgments, headless instead of GUI automation, packaged as a reusable SKILL.md — converging on code-as-step-between-model-runs, agent-native-computing, and skills-and-workflows. It stays LOW because it is domain transfer, not bearing: it challenges and extends nothing, leaves the natural code-path-vs-MCP contrast (Willison's raw --background --python vs the independents' MCP servers) unadjudicated where why-code-execution-beats-mcp would care, and the open scene-editability question goes unanswered — a dated receipt for the canon, not news that would change what he builds or argues.
ip:concept.code-as-step-between-model-runsip:framework.agent-native-computingip:framework.code-first-architectureip:concept.skills-and-workflowsip:source.why-code-execution-beats-mcpradar:concept.mcpradar:concept.desktop-agentsradar:concept.agent-skillsradar:nomoreda-agent-facing-eda
queries asked of Scott's wikis
- code as the step between model runs editable artifacts
- agent skill files SKILL.md reusable markdown skills
- GUI automation versus programmatic CLI agent control
- MCP server integrations coding agents local tools
- agent feedback loop inspect render screenshot revise
- agent-operable website local app headless execution
| source | object | author | score | comments |
| 🟧 hn | Using Blender with coding agents on macOSRetrieved article excerptOpen article · Retrieved 2026-09-17T07:22:37.158760+00:00 # Using Blender with coding agents on macOS
Modern frontier models have got *really good* at using Blender. I've been having a lot of fun trying this out recently - models can produce `.blend` files you can edit in Blender itself, and can also render images and even movies (by rendering a sequence of images and combining them with `ffmpeg`).
Setting this up, at least on macOS, is *really easy*. Install the Blender desktop app from [blender.org](https://www.blender.org) and then tell the coding agent:
> `Use the already install /Applications/Blender to render a scene of a pelican riding a bicycle`
That worked for me with GPT-6 Astra. You can also be a bit more explicit, to save the model some time figuring out how to use it:
> `Use Blender like this: /Applications/Blender.app/Contents/MacOS/Blender --background --python scene.py`
## A pelican riding a bicycle with GPT-6 Astra
I used GPT-6 Astra (Medium) in the ChatGPT macOS application, Codex mode:
> `Use the already install /Applications/Blender to render a scene of a pelican riding a bicycle`
2m39s later:
[A whimsical 3D illustration of a white pelican riding a turquoise bicycle with cream mudguards and silver spokes. Its wings stretch forward to grip the handlebars, and its long yellow legs bend down to the pedals. A coral scarf streams behind its neck. The pelican has a large glossy eye and a long pale yellow bill with a rounded throat pouch. Soft lighting and a muted sage-green background give the scene a pastel, toy-like appearance.](https://raw.githubusercontent.com/simonw/gpt-6-astra-blender-pelican-bicycle/refs/heads/main/outputs/pelican-bicycle.png)
[pelican-bicycle.blend](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle/blob/main/outputs/pelican-bicycle.blend), [pelican\_scene.py](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle/blob/main/work/pelican_scene.py)
> `OK add a background and a lot of flair`
3m51s later:
[A whimsical pastel 3D illustration of a white pelican riding a turquoise bicycle along a pink boardwalk. It wears a cream boater hat with a pink band and a coral scarf fluttering behind it, with wings reaching for the handlebars and long yellow legs on the pedals. Three balloons trail behind, and a colorful pinwheel is mounted near the front wheel. Striped beach huts, pink-and-white parasols, and angular palm trees line the beach beneath a teal sky. Pastel triangular bunting hangs overhead, while confetti and curling streamers fill the festive scene.](https://raw.githubusercontent.com/simonw/gpt-6-astra-blender-pelican-bicycle/refs/heads/main/outputs/pelican-bicycle-festival.png)
[pelican-bicycle-festival.blend](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle/blob/main/outputs/pelican-bicycle-festival.blend), [pelican\_flair.py](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle/blob/main/work/pelican_flair.py)
> `OK make it a whole lot better`
5m59s later:
[A whimsical 3D illustration of a white pelican cycling along a seaside boardwalk at sunset. It wears a cream boater hat and a coral scarf, with wings on the handlebars and long orange legs reaching the pedals of a turquoise bicycle. A wicker front basket holds pink and white flowers, and three balloons float behind. Pastel bunting stretches overhead between palm trees. Striped beach huts stand beside a teal sea with a small sailboat, beneath a large peach-colored sun. The scene has a softly lit, toy-like style.](https://raw.githubusercontent.com/simonw/gpt-6-astra-blender-pelican-bicycle/refs/heads/main/outputs/pelican-coastal-parade.png)
[pelican-coastal-parade.blend](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle/blob/main/outputs/pelican-coastal-parade.blend), [pelican\_final.py](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle/blob/main/work/pelican_final.py)
The project lives [in this repo](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle), and [here's the exported transcript from Codex](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle/blob/main/codex-transcript.md).
## Creating a skill
Codex makes it pretty easy to create skills (using its built-in skill creating skill), so I finished up by prompting:
> `Create a quick skill that describes how to use the currently installed /Application/Blender based on what you learned`
It produced and installed [this Markdown skill](https://github.com/simonw/gpt-6-astra-blender-pelican-bicycle/blob/main/outputs/blender-local/SKILL.md), which I have since used for further Blender experiments with prompts like this:
> `Use your Blender Local skill to build this scene (attached image)`
Created 2026-09-05T08:42:28-07:00, updated 2026-09-05T08:47:13-07:00 · [History](https://github.com/simonw/til/commits/main/llms/blender-coding-agents-macos.md) · [Edit](https://github.com/simonw/til/blob/main/llms/blender-coding-agents-macos.md) | activatedgeek | 6 | 1 |
| 🟧 echo.blog ⭐ | Willison reports generating and revising a pelican-on-a-bicycle scene with GPT-6 Astra in Codex mode, publishing the Python scripts, editabl | Simon Willison | — | — |
| 🟠 reddit | Astra/Codex working on Blender 3.6.23 through my MCP backport. Looking for testers. OpenAI | SpinachOk9137 | 0 | 0 |
| 🟠 reddit | Anyone combined GPT-6 Astra + Higgsfield AI in Blender via MCP to save tokens for 3D printing? artificial | Julianw20012 | 0 | 4 |
| 🟠 reddit | Retopologizing facial mesh with Codex + Astra + Blender OpenAI | nmfisher | 2 | 1 |
| 🟧 hn | Show HN: Blender Copilot | xeonax | 6 | 0 |
| 🟧 hn | Show HN: Maniac Mansion Reconstructed in 3D from Its Original Screens | s-macke | 2 | 0 |
| 🟧 hn | Show HN: A 3D bamboo grove inspired by my childhood home | YuukiJyoudai | 1 | 0 |
2026-09-28T17:11:53Z
The bamboo grove is the seventh independent implementation and second photo-to-3D reconstruction, but it confirms rather than extends: Willison's claim has proved out into an established pattern — agents write Python that drives Blender headless, via MCP, harnesses, or direct CLI — so the case absorbs. Heat drops not because the periphery stopped spreading (three implementations landed in the final 72h) but because the case's question is answered and the continuing diffusion already lives under the hot coding-agents/agent-harnesses concept tracks; keeping it open would tax Scott's attention for receipts his canon already holds.
2026-09-28T15:45:05Z
evidence attached: hn.story.49878073 — Independent builder used GPT-6 Astra to turn photos into an editable Three.js 3D scene — organic corroboration of the Astra-driven 3D-workflow episode's spread.
2026-09-27T09:50:24Z
s-macke's Show HN extends the pattern to image→3D reconstruction via iterative Astra+Blender scripting — a chair refined through a visible render-inspect-revise loop (chair-evolution.gif), then a full Maniac Mansion rebuild from original screens, code released, invocation mechanism unstated. That makes six independent implementations in ~3 weeks across three platforms and five domains, and the second apparent non-MCP Astra+Blender instance after Willison; with arrivals still steady while engagement stays dead, the pattern itself — not the numbers — now carries the case into accelerating, with heat held medium on the continuing periphery.
2026-09-27T09:23:29Z
evidence attached: hn.story.49864738 — Independent Astra+Blender artifact (image-to-3D scene reconstruction via iterative scripting) corroborates the emerging Astra-driven Blender workflow pattern from a second author with released code.
2026-09-26T18:54:21Z
The newest independent implementation (Show HN 'Blender Copilot', 2 pts/0 comments) shifts the pattern's meaning again: LLM-driven Blender editing is now model-agnostic and self-tooling — a DeepSeek-generated in-viewport Python harness built in a day for ~$7 — moving from one-off Astra demonstrations toward standing tools, while still never rerunning Willison's background-Python path. The numbers are dead (0.17 pts/h vs 20.5 peak, 21 days old), but three implementations arrived in ten days across distinct communities and domains, so heat holds medium on the expanding periphery, not the engagement.
2026-09-26T18:25:02Z
evidence attached: hn.story.49859013 — An independent in-viewport Blender harness editing the live scene via LLM-written Python corroborates the spread of the agent-edits-Blender-through-code pattern beyond Willison's workflow.
2026-09-25T01:53:26Z
grounded: converges/low — Willison plus four independent practitioners keep newly arriving at a position Scott's canon already holds — the model writes Python that Blender executes betwe
2026-09-25T01:47:38Z
A fourth independent practitioner (nmfisher) demonstrates Astra+Codex retopologizing a facial mesh in Blender, head-to-head against DeepSeek-V4.1-Flash — Astra clearly better though neither great, both completing in under an hour with loose instructions, video published. This extends the pattern from scene generation and fabrication prep to technical mesh editing and adds the first comparative model signal, which promotes the broad pattern to corroborated; the exact macOS background-Python path remains unreproduced (all independents use MCP), and engagement is negligible, so heat holds on the expanding periphery rather than the numbers.
2026-09-25T01:26:13Z
evidence attached: reddit.post.1wpj1o4 — Independent user's Astra-driven Blender retopology test corroborates the emerging pattern of programmatic Astra 3D-artifact editing from the Willison case.
2026-09-19T10:22:20Z
A new independent user reports roughly 22 Blender models in a multicolor 3D-printing workflow, extending the evidence from illustrative scenes toward downstream fabrication. This modestly strengthens practical utility, but the truncated MCP-based account neither reproduces Willison’s background-Python path nor verifies print success or sustained editability.
2026-09-19T10:21:52Z
evidence attached: reddit.post.1wki3h0 — This is independent usage evidence that Astra plus Blender MCP can produce practical editable 3D artifacts, directly bearing on the open workflow case.
2026-09-17T18:35:28Z
A separate developer reports successful Astra/Codex scene generation through a Blender 3.6.23 MCP backport, modestly strengthening the broader practicality claim. The truncated report does not independently reproduce Willison’s background-Python workflow or establish that generated scenes remain well structured for subsequent editing.
2026-09-17T18:22:33Z
evidence attached: reddit.post.1wj0jn0 — A separate working MCP backport and substantial Blender artifact provide useful corroboration that Astra-driven editable 3D workflows are practical.
2026-09-17T07:28:40Z
grounded: converges/low — Willison’s reported script-driven Blender workflow converges with Scott’s Code as the Step Between Model Runs concept and his agent-operable website stack’s pre
2026-09-17T07:25:55Z
case created — Published scripts, editable outputs, and a transcript make this a bounded reproducible workflow claim rather than generic enthusiasm about agents using creative tools.