Inspiration
Blender is extraordinarily powerful, but a beginner must understand coordinates, object hierarchies, materials, lighting, cameras, animation, rendering, and recovery before a visual idea becomes a finished scene. Codex3D lets a creator describe that idea in ordinary language, watch it appear progressively in Blender, refine it through follow-up feedback, and safely reject a bad direction.
What it does
Codex3D is a local-first natural-language 3D agent for Blender. Codex with GPT-5.6 translates creative intent into typed MCP tools. An authenticated localhost bridge queues structured actions for Blender's main thread, then returns compact scene inspection, render validation, and snapshot results for the next turn.
The demonstrated workflow includes:
- Progressive creation of geometry, cyclic 3D curves, materials, lights, camera, world, and animation
- Multi-turn visual follow-up without breaking an approved six-curve structure
- Stable object UUIDs across rename and .blend reload
- Session-bound snapshots with an automatic pre-restore safety snapshot
- Fingerprint-verified restore after a rejected creative edit
- Composition checks for crop, centering, unwanted objects, curve visibility, luminance, and overexposure
- Versioned .blend, validated PNG preview, and H.264 MP4 delivery
How it works
User feedback
-> Codex with GPT-5.6
-> local STDIO MCP adapter
-> authenticated 127.0.0.1 bridge
-> bounded Blender main-thread pump
-> structured bpy backend
-> inspection / validation / snapshot result
-> next feedback turn
Codex3D exposes twenty-three allowlisted MCP tools and never exposes arbitrary Python or shell execution. The MCP and API layers do not import bpy, and the socket worker never mutates Blender directly.
How Codex and GPT-5.6 helped
Codex collaborated throughout the build: protocol design, the installable Blender extension, localhost transport, typed MCP schemas, fake and real backends, tests, isolated acceptance runners, release audit, and documentation. GPT-5.6 was used meaningfully to reason across protocol, transport, Blender lifecycle, safety, visual validation, and multi-turn recovery boundaries.
Key human-directed decisions were Blender-first local execution, structured tools instead of generated scripts, localhost-only authentication, artifact confinement, stable identity, reversible recovery, and honest separation of synthetic evidence from human usability evidence.
Challenges
- Blender scene access must happen on its main thread while Codex runs in another process.
- Initial unattended approval semantics cancelled restore and checkpoint delivery.
- Basic nonblack render checks missed cropped, off-center, overexposed, and unwanted objects.
- Large scene inspections polluted long multi-turn context.
Accomplishments
- 177 passing automated tests in the latest closeout regression
- Unchanged nine-turn synthetic novice-language test: SYNTHETIC_PASS 100/100
- Real Blender 5.1.1 snapshot restore, persistent UUID, preview, and animation verification
- Compact inspection reduced response size by 91.1% in the repaired test
- Public MIT-licensed repository and prebuilt Blender Extension release
Testing
The repository README contains a 60-second quick start, Blender Extension installation, MCP setup, and a no-Blender fake-backend judge path. Judges can use the prebuilt release without rebuilding the extension from scratch.
Repository: https://github.com/kytianhong/codex3d Release: https://github.com/kytianhong/codex3d/releases/tag/hackathon-v0.1.0 Demo: https://youtu.be/symobXoyBKA
Limitations
The full nine-turn synthetic workflow takes about 17.5 minutes. Broad mutation batches may require approval. There is no Web UI and full Scene IR reconciliation is not implemented. The synthetic test demonstrates repeatable functionality, not human usability research.
Codex3D is an unofficial Codex-inspired concept demo. It is not endorsed by OpenAI and does not reproduce an official logo.
Log in or sign up for Devpost to join the conversation.