Important
This README is a quick skim. Full documentation lives in docs/README.md — bookmark that page for everything (specs, runtime, install, evals, laws, skills, recipes, research).
- A portable AI-agent Skills pack for Blender, built for users who don't know Blender.
- A content layer (
SKILL.md-based) that loads inside any compatible AI agent. - Complements Anthropic Blender Connector + Blender Lab MCP — does not replace them.
Human intent (non-Blender user)
→ BlendOps Skills (planning · readiness · evidence · handoff)
→ external Blender runtime (Path 1 Lab MCP / Path 2 community / CLI)
→ validated artifacts + non-Blender-user response| BlendOps publishes (this repo's deliverable) | BlendOps redirects to (upstream — not this repo's work) |
|---|---|
| 48 skills (baseline + web 3D deep + quality validators + domain planners + process discipline + asset/sourcing + color/format/output + recipe/pack ops + runtime path discipline, …) | Path 1 — Blender Lab MCP (Lab add-on + Lab server). Owner: Blender Foundation. Source: blender.org/lab/mcp-server |
| 4 laws (evidence-before-done, official-runtime-only, no-arbitrary-python-interface, non-blender-user-language) | Path 1 host (a) — Anthropic Blender Connector (Claude-Desktop-specific MCP host). Owner: Anthropic. Source: claude.com/.../using-the-blender-connector-in-claude |
1 pack (product-hero-v0), 3 bundles (skill-package, claude-desktop-manual, generic-project-local) |
Path 2 — Community mcp-for-blender (formerly blender-mcp) (different add-on/server). Owner: mature community project. Source: github.com/ahujasid/mcp-for-blender |
26 install target docs + 35+ CLI targets (--with-skills-dir, skills.sh.json, skills-lock.json v2) + 3 mode auto-detect install flow |
CLI fallback (blender --background --python ...). Owner: Blender Foundation. Source: docs.blender.org/manual/.../command_line |
Skill ZIP / OpenAI YAML / canonical bundle generation (npm run skills:export → 48 ZIPs, one per skills/*/SKILL.md) + distributable local CLI (bin/blendops.mjs / scripts/install-skills.mjs, packable via npm pack) + GLB 2.0 web handoff toolkit (blendops glb-inspect, blendops glb-handoff, blendops glb-validate-manifest, Three.js & R3F integration examples in docs/examples/glb-web-handoff/README.md) |
The Blender executable, MCP servers, MCP host clients, OS, GPU drivers, network. |
What this means for "completeness":
- The skills layer (48 skills + 151 deep reference guides across 8
skills.sh.jsongroups + 26 install docs + 35+ CLI targets + ZIP/package export + deterministic CI guards) is complete and production-ready atv1.0.0— nothing upstream blocks it. Milestones:v0.4.0-draft(40 skills) →v0.5.0-draft(48 skills) →v0.8.0-draft(hardening) →v0.9.0-draft(catalog & CLI readiness) →v1.0.0(production release). - Whether the upstream runtime paths work in your environment is up to upstream Blender Foundation / Anthropic / community
mcp-for-blender(Path 1 + Path 2 are user-reported verified by the repo owner — operator statement, does not promote eval scope; CLI is upstream-documented). BlendOps redirects to them via per-target install docs but does not own their stability. - The "Full runtime eval: Not Run" / "Artifacts: Not Produced" labels refer to in-repo provenance evidence files only — operator-side records under
docs/evals/. They do not mean "the upstream runtime paths don't work" or "BlendOps is incomplete."
Start at blendops-help or jump straight into the narrowest group from skills.sh.json and skills/README.md:
Goal unclear or messy request? ──▶ blendops-help / blender-brainstorming ──▶ intent-to-3d-brief-writer
Know what you want to build? ──▶ recipe-fit-assessor / workflow-stage-router
├─ Product / Catalog ──▶ product-hero-scene-planner / product-grid-scene-planner
├─ Character / Env ──▶ character-portrait / environment-shot / interior-scene
└─ Complex (4+ steps)──▶ blender-recipe-decomposer ──▶ stage-gates ──▶ done-criteria
Need visual / budget checks? ──▶ lighting-quality-checker • material-quality-checker • composition-quality-checker • polycount-budget-checker
Need render / format rules? ──▶ color-management-decision • cycles-vs-eevee-decision • resolution-aspect-decision • output-format-decision
Ready for Web 3D / GLB? ──▶ mobile-glb-budget ──▶ glb-animation-handoff / r3f-component-plan ──▶ glb-web-handoff
Setting up or auditing runtime? ──▶ runtime-path-picker ──▶ runtime-readiness ──▶ runtime-attempt-report (or bridge-conflicts)
Ready to deliver / claim Pass? ──▶ scene-readiness ──▶ render-export-evidence ──▶ pre-handoff-verification ──▶ plain-language-responsePaste this into any AI agent:
Read and follow the BlendOps AI Agent Quickstart:
https://raw.githubusercontent.com/ThanhNguyxnOrg/blendops/main/docs/ai-agent-quickstart.md
Auto-detect your environment first, then pick one mode:
- If you have project filesystem write access → Mode A: install BlendOps project-locally (skills/, laws/, packs/, target-specific entrypoint).
- If you are in a chat-only UI (Claude Desktop chat, Claude.ai, ChatGPT) with no filesystem write → Mode B: output 48 separate ZIPs from skills/*/SKILL.md (one ZIP per skill) for me to drag-drop into the Skills UI manually.
- If you cannot fetch the repo, cannot write, AND cannot output ZIPs → Mode C: ask one concise question.
Do not install Blender, configure runtime connectors, run Blender, or claim generated artifacts. Report the selected mode, files changed or ZIP filenames, runtime status, artifact status, and limitations.The BlendOps GitHub repository is already listed on skills.sh. Install the skill collection with the Agent Skills CLI:
npx skills add ThanhNguyxnOrg/blendopsThe Skills CLI guides you through choosing skills and supported agent targets. The repository's skills.sh.json defines eight display groups for the listing. It is source metadata; no ZIP upload or separate npm publication is required for skills.sh.
The public npm package is blendops@1.0.0 (dist-tag latest). No clone or global install is required:
npx blendops --help
npx blendops doctor --project .
npx blendops install --target auto --project . --dry-run
npx blendops install --target auto --project . --writeReview the dry-run plan before running the matching --write. The installer changes only the selected project; it does not install Blender or configure an MCP runtime.
For contributors working from a local clone, use the repository scripts:
npm run skills:install -- --target universal --project /path/to/your-project --with-skills-dir --dry-run
npm run skills:install -- --target universal --project /path/to/your-project --with-skills-dir --writeThe agent auto-detects which mode applies. Coding agents (Claude Code, Cursor, Codex, OpenCode, Cline, Roo Code, Blackbox AI, Continue, Zed, Windsurf, Trae, Augment Code, goose, OpenHands, JetBrains Junie, Sourcegraph Cody, Aider, Amazon Q Developer, Gemini, Copilot, Antigravity, Kiro, Amp, Droid, Kilo, OpenClaw, ClawdBot, VS Code, universal .agents/skills/) hit Mode A and install project-locally. Claude Desktop / Claude.ai / ChatGPT chats hit Mode B and output separate ZIPs from skills/*/SKILL.md (one per skill, 48 total) for you to drag-drop into the Skills UI manually. Anything else falls back to Mode C — blocked-needs-input. Full mode logic in docs/ai-agent-install-flow.md.
BlendOps recognizes 2 MCP execution paths + a CLI fallback appendix (full doc):
- Path 1 — Official Blender Lab MCP (Lab add-on + Lab server installed in Blender 5.1+, hosted from either (a) Anthropic Blender Connector in Claude Desktop, or (b) any other MCP client configured manually).
- Path 2 — Community
mcp-for-blender(formerlyblender-mcp) (different add-on/server, mature third-party, Blender 3.0+). - CLI fallback (appendix) — direct
blender --background --python. Documented upstream as a first-class Blender CLI surface; no in-repo evidence file yet.
Warning
Anthropic Connector is not standalone. Anthropic's tutorial step 2 explicitly tells you to install the Blender Lab MCP add-on inside Blender. The Connector is the Claude-Desktop-specific MCP host on top of the Lab stack. Path 1 needs Blender 5.1+ regardless of host because of the Lab add-on manifest.
| Want to | Open |
|---|---|
| Read everything (the hub) | docs/README.md |
| 5-minute orientation | docs/start-here.md |
| Set up Blender runtime | docs/external-runtime-setup.md |
| Install BlendOps in your AI tool | docs/install/README.md |
| Inspect a GLB & generate a web handoff report (Three.js / R3F) | docs/examples/glb-web-handoff/README.md |
| Browse skills / laws / packs | skills/README.md · laws/README.md · packs/README.md |
| Check release readiness | docs/release-readiness.md |
| Inspect repository architecture & file inventory | docs/architecture.md · docs/audits/current/repository-inventory.json |
| See the canonical portable skill package | bundles/skill-package/blendops/ |
BlendOps does not install Blender, ship its own runtime, configure Anthropic Connector or the Lab MCP add-on for you, run Blender, claim mutation/render/export artifacts, or claim an official marketplace/plugin listing. npm and skills.sh provide content/CLI distribution only; Blender runtime setup is separate. See docs/release-readiness.md for the honest current state.
MIT — see LICENSE.