by jangtrinh
AI-agent operating system for Blender 5.2: skills, verified bpy knowledge base, AGENT_OK/AGENT_FAIL execution contract, production gate for 3D-printable parts, robot-arm demo
# Add to your Claude Code skills
git clone https://github.com/jangtrinh/design-os-3d-blenderGuides for using ai agents skills like design-os-3d-blender.
design-os-3d-blender is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by jangtrinh. AI-agent operating system for Blender 5.2: skills, verified bpy knowledge base, AGENT_OK/AGENT_FAIL execution contract, production gate for 3D-printable parts, robot-arm demo. It has 55 GitHub stars.
design-os-3d-blender's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/jangtrinh/design-os-3d-blender" and add it to your Claude Code skills directory (see the Installation section above).
design-os-3d-blender is primarily written in Python. It is open-source under jangtrinh on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh design-os-3d-blender against similar tools.
No comments yet. Be the first to share your thoughts!
Unlocks once the catalog security scan passes (runs nightly).
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
Contents: Product tour · What is inside · Proof · ORC workflow demo · Quick start · The loop · Worked examples · Notes · License
An AI-agent operating system for Blender 5.2 LTS: skills, a verified knowledge base, an execution contract, and a production gate — built so an agent (Claude Code, Codex, Google Antigravity, or any MCP client) can model, rig, animate, render, and ship 3D-printable, dimension-correct parts rather than pretty demos.
Everything here is Blender-native: no vendor mesh generation, no downloaded assets, no paid model calls.
The 24-second loop pairs overview and detail frames from four native Blender builds: robot arm, watch-winder capsule, wellhead, and separator. Frame labels state the evidence scope; it is not a physical, manufacturing, pressure, thermal, or performance claim. Static fallback poster · Source and output provenance.
| Layer | Path | Purpose |
|---|---|---|
| Operating rules | AGENTS.md (canonical, Codex) · CLAUDE.md (imports it) · .project-agent.md (binding rules) |
Contract-first loop, sentinel execution contract, verify ladder, failure map from real sessions |
| Skills | .agents/skills/blender-agent-core, blender-image-to-3d, blender-knowledge-workbench (.claude/skills mirrors) |
Router + recipes; reference-image reconstruction; catalog-driven reading packs |
| Knowledge base | knowledge/ (28 domain files, 5.2-verified) + research/ (engineering syntheses) |
Data-API-first bpy patterns, decision tables, failure→fix tables, standards for CAD/print/robotics |
| Execution | scripts/agent_runtime.py, scripts/headless-run.sh, scripts/blender-socket-client.py |
Every payload ends with AGENT_OK {json} / AGENT_FAIL {json} (full traceback), exit 0/1/2/3 — Blender's own exit code is unreliable both ways |
| Verification | scripts/agent-verify-lib.py (+ scripts/agent_verify/) |
Numeric "sense organs": framing, frame stats, state-restoring previews, isolated export re-import |
| Production gate | specs/build-spec.schema.json, scripts/production-gate.py (+ scripts/production_gate/) |
Spec → manifold/self-intersection/volume, dimensions ± tolerance, wall/overhang screens, hole diameters vs ISO 273 / ISO 4762 / heat-set tables, STL round-trip + sha256 manifest; report bound to scene+spec hashes |
| Boilerplates | scripts/boilerplates/ (30 modules) |
CAD, gears, fasteners, O-rings, connectors, rigs, geometry nodes, render — each self-tests under the sentinel contract |
| Tests | tests/execution, tests/production-gate, tests/knowledge |
Execution-contract, production-gate, and knowledge-catalog suites |
| MakerWorld leg | makerworld-pipeline/ (validator scripts/validate_bambu_3mf.py, fixture generator tools/, docs) |
Validates a Bambu Studio project 3MF (geometry through the production extension, per-plate bbox, bed fit, thumbnails, settings); sourced publishing constraints and a human publish checklist — printing and publishing stay manual by MakerWorld policy |
| Agent runtimes | AGENTS.md (Codex, any AGENTS.md reader) · CLAUDE.md (Claude Code) · .agents/rules, .agents/workflows, .agents/mcp_config.json (Google Antigravity) |
Same skills folder (.agents/skills) is read natively by Antigravity; setup in docs/antigravity-setup.md |
| Worked examples | builds/robot-arm-*, builds/watch-winder-capsule |
Two builds: the robot-arm session (accepted assembly film + 38-part print kit) and the watch-winder capsule (native render asset + gate-passed print plates) — each with its honest limits |
![]() |
![]() |
The model open in Blender 5.2.0 — scene A5-Original-refined, 671 objects, frame 3031/3031, outliner and properties visible (full-window screenshot, not a viewport crop) |
Six-angle Cycles orbit of the assembled parts at the final frame, rendered headless from the shipped .blend |
![]() |
![]() |
| Render tessellation of the 402 arm part meshes (114,723 polygons), Cycles Wireframe node — every edge you see is in the file | 38 fit-prototype parts arranged on five 220×220 mm plates (PETG + TPU), exported as STL/3MF |

Cycles render of the assembled model at the final frame (1600×1200, 96 samples, Metal). The full-speed MP4 of the assembly film is in builds/robot-arm-original-refined/video/.
Production gate run on the real print kit (2026-09-06). The 38 exported STLs were re-imported into a metre-scaled scene and gated against a spec generated from parts-list.csv (target dimensions ± 0.1 mm, one shell per part, PETG/TPU):
GATE PASS | 38 part(s), 0 failing check(s) | scene=72ab501e47bf spec=d88bc836fb3a | Blender 5.2.0 LTS
836 checks pass · 0 fail · 38 skip (no min_wall_mm declared) · 38 info (overhang reported, not gated)
Per part: non_manifold_edges, non_contiguous_edges, wire_edges, loose_verts, zero_area_faces, self_intersection_pairs, shells, signed_volume_positive, bbox_dims_mm, scale_applied, scene_unit_system, scene_scale_length, then STL round-trip repeats of the topology and dimension checks. Report, spec and the reproduction script: builds/robot-arm-print-assembly/reports/gate-report-260906.json, gate-spec-260906.json, scripts/gate-import-parts.py. The report's exclusions say what this does not prove: load capacity, print success, assembly fit, thermal/creep — those remain physical evidence, and the arm stays unreleased for manufacture.
The ORC component workflow, brief template, and Separator worked example show how a reference-driven component moves from coverage and interface contracts to object-level evidence and bounded review.
The 24-second component tour uses native review media and bounded family studies. Wellhead and separator owner appearance review remains pending; bellows R2 is a form-approved reusable pilot; the RTD/PT images are form studies, not a final exchanger-train claim. Static fallback poster · Source and output provenance.
![]() |
![]() |
| Completed render-only Separator component: overview. | Completed render-only Separator component: a critical crop used to inspect the vapor nozzle, flange, and attached instrument. |
Form-approved reusable bellows-family pilot. Exchanger-train detail, a full plant, print readiness, and manufacturing qualification remain out of scope.
The three images are compressed derivatives of native Blender renders, not source references. Provenance records the source and output hashes plus the evidence scope.
Requirements: macOS/Linux, Blender 5.2.x, Python 3.1