v0.7.0-draft: +6 MCP-aware coding agents + Antigravity upgrade (26 install targets)

BlendOps v0.7.0-draft — +6 MCP-aware coding agents + Antigravity upgrade (26 install targets)

Adds 6 popular AI coding agents that ship native MCP support, plus upgrades Google Antigravity from "not researched" to a fully linked-only install doc. Skills layer remains 48 / 48 (= 3× baseline).

What changed since v0.6.0-draft

+6 new install targets (all native MCP, web-research verified)

• Blackbox AI (VS Code) — Blackbox panel → MCP Servers. Source: docs.blackbox.chat/.../mcp-servers.
• Augment Code — Easy MCP one-click + Settings panel + ~/.augment/settings.json (http / sse / stdio + ${workspaceFolder} expansion). Source: docs.augmentcode.com/setup-augment/mcp.
• Aider (terminal) — aider --mcp-servers '{...}' flag or .aider.conf.yml; LiteLLM-backed; stdio only, 25 tool calls per reply cap. Source: PR Aider-AI/aider#3937.
• Trae (ByteDance) — Settings → MCP → Add Manually (v1.3.0+); supports local stdio + remote HTTP servers. Source: docs.trae.ai/ide/add-mcp-servers.
• Sourcegraph Cody — agentic context fetching with MCP tools; opt-in; Cody decides when to invoke. Source: sourcegraph.com/docs/cody/.../agentic-context-fetching.
• Amazon Q Developer CLI — qchat mcp add/remove/list/import/status; agent config under ~/.aws/amazonq/cli-agents; local + remote (HTTP/OAuth) MCP. Source: docs.aws.amazon.com/.../command-line-mcp-config-CLI.html.

Antigravity upgraded "not researched" → "linked-only"

Google Antigravity now ships a fully written install doc covering MCP Store GUI + manual mcp_config.json flow, including upstream prerequisites (uv, Node.js, IAM roles for Data Agent Kit). Source: codelabs.developers.google.com/google-workspace-mcp-antigravity.

Docs sync — 20 → 26 install targets

Every active-claim doc updated:

• README.md (badge Install targets: 26, scope table, agent enumeration in 30-second prompt, +7 agent badges)
• docs/README.md install hub (1 chat UI + 21 coding agents + 3 local LLM runners + 1 generic)
• docs/install/README.md, docs/ai-agent-install-flow.md, docs/multi-agent-install-strategy.md
• docs/release-readiness.md (dashboard + Ready list)
• AGENTS.md, package.json, TODO.md
• scripts/check-docs.mjs requiredDocs

CI

• npm run docs:check validates 352 active markdown files (+6 new install docs).
• npm run skills:export deterministically produces 48 ZIPs + 48 Claude Code folders.

Full milestone history

Tag	Skills	Install targets	Theme
v0.1.0-draft	16	9	Baseline
v0.2.0-draft	24	16	+ Web 3D + Quality validators
v0.3.0-draft	32	16	+ Domain planners + Process discipline
v0.4.0-draft	40	16	+ Asset/sourcing + Color/format/output
v0.5.0-draft	48	16	+ Recipe/pack ops + Runtime path discipline
v0.6.0-draft	48	20	+ Roo Code, Windsurf, OpenHands, JetBrains Junie
v0.7.0-draft	48	26	+ Blackbox AI, Augment Code, Aider, Trae, Cody, Amazon Q + Antigravity upgrade

Assets (48 portable Skill ZIPs)

Drag-drop directly into Claude.ai / Claude Desktop / ChatGPT Skills UI, or extract into .claude/skills/ for Claude Code, or wire into any of the 26 install targets documented in docs/install/.

Scope reminder (unchanged)

BlendOps still publishes one thing: the skills layer. Runtime paths (Path 1 Blender Lab MCP, Path 2 community ahujasid/blender-mcp, CLI fallback) remain upstream's deliverable — see docs/runtime-stack-strategy.md. Connector is a Path 1 host, never standalone.

v0.6.0-draft: final agent expansion + docs sync (48 skills + 20 install targets)

BlendOps v0.6.0-draft — final agent expansion + docs sync

Final Draft v0 release. The skills layer is locked at 48 skills (3× baseline), the install matrix now covers 20 targets, and every active-claim doc is synced to those numbers.

What changed since v0.5.0-draft

+4 install targets (popular MCP-aware coding agents)

All four are linked-only confidence (matching the existing pattern for newer agents) — BlendOps stays project-local-first; runtime is the upstream client's job.

• Roo Code (VS Code) — .roo/mcp.json (project) or global mcp_settings.json. Source: docs.roocode.com/features/mcp/using-mcp-in-roo.
• Windsurf (Codeium Cascade) — ~/.codeium/windsurf/mcp_config.json; 100-tool ceiling. Source: docs.windsurf.com/windsurf/cascade/mcp.
• OpenHands (formerly OpenDevin) — ~/.openhands/mcp.json + Settings UI + openhands mcp CLI. Source: docs.openhands.dev/openhands/usage/settings/mcp-settings.
• JetBrains Junie / AI Assistant — Settings → Tools → Junie → MCP Settings, or .junie/mcp/mcp.json (project) / ~/.junie/mcp/mcp.json (user). Source: junie.jetbrains.com/docs/junie-plugin-mcp-settings.html.

Docs sync — 16/40 → 48 skills + 20 install targets

Every active-claim document now agrees: 48 skills, 20 install targets, "one ZIP per skills/*/SKILL.md". Stale phrasings like "10 ZIPs", "9 targets", "16 ZIPs" cleaned up across:

• README.md (badge, scope table, 30-second prompt)
• docs/README.md install hub (1 chat UI + 15 coding agents + 3 local LLM runners + 1 generic)
• docs/install/README.md, docs/ai-agent-install-flow.md, docs/multi-agent-install-strategy.md
• docs/release-readiness.md (dashboard + Ready list)
• docs/ai-agent-quickstart.md Mode B
• bundles/claude-desktop-manual/BLENDOPS_SINGLE_FILE.md (baseline 16 + expansion routes)
• skills/blendops-help/references/lifecycle-phases.md
• AGENTS.md, CONTRIBUTING.md, .github/pull_request_template.md, package.json, TODO.md

Historical phase logs (e.g. "Phase 2.6: 10 skills total", "Phase 4: 16 skills now Expanded Draft v0") are intentionally preserved as snapshots of what was true at the time.

CI

• npm run docs:check validates 346 active markdown files (was 342 — +4 new install docs).
• npm run skills:export produces 48 ZIPs + 48 Claude Code folders deterministically; scripts/check-docs.mjs registers all 4 new install docs.

Full milestone history

Tag	Skills	Install targets	Theme
v0.1.0-draft	16	9	Baseline
v0.2.0-draft	24	16	+ Web 3D handoff deep + Quality validators
v0.3.0-draft	32	16	+ Domain scene planners + Process discipline
v0.4.0-draft	40	16	+ Asset/sourcing + Color/format/output
v0.5.0-draft	48	16	+ Recipe/pack ops + Runtime path discipline
v0.6.0-draft	48	20	+ Roo Code, Windsurf, OpenHands, JetBrains Junie + final docs sync

Assets (48 portable Skill ZIPs)

Drag-drop directly into Claude.ai / Claude Desktop / ChatGPT Skills UI, or extract into .claude/skills/ for Claude Code, or wire into any of the 20 install targets documented in docs/install/.

Scope reminder (unchanged)

BlendOps still publishes one thing: the skills layer. Runtime paths (Path 1 Blender Lab MCP, Path 2 community ahujasid/blender-mcp, CLI fallback) remain upstream's deliverable — see docs/runtime-stack-strategy.md. Connector is a Path 1 host, never standalone.

v0.5.0-draft: 48 skills (3x baseline reached)

BlendOps v0.5.0-draft — 48 skills (3x baseline target reached)

Final Draft v0 milestone for the BlendOps skills layer roadmap. The repo now ships 48 portable Anthropic Skills = 3x the original 16-skill baseline.

What changed since v0.4.0-draft

Batch 7 — Recipe / pack operations (4 skills)

• recipe-fit-assessor — map user goals to BlendOps recipes/packs; flag scope gaps; recommend canonical workflow entry
• pack-prerequisite-checker — GO/NO-GO matrix (brief, assets, runtime posture, evidence) before pack execution
• workflow-stage-router — pick the next narrow skill from current artifact inventory + preconditions
• acceptance-criteria-builder — turn locked brief + recipe into measurable checklist with evidence hooks

Batch 8 — Runtime path discipline (4 skills)

• runtime-path-picker — Path 1 Lab MCP vs Path 2 ahujasid vs CLI appendix decision under constraints
• path-one-host-outline — Anthropic Connector vs manual MCP host; Lab add-on inside Blender 5.1+ always required
• cli-appendix-job-outline — phased headless blender --background job outline; upstream CLI manual is authority
• runtime-attempt-report-writer — truth-label evidence-bound attempt report after Path 1 / Path 2 / CLI tries

Each new skill ships with SKILL.md + EVAL.md + 3 references/ files (32 new reference files).

Roadmap status — 48/48 (100%)

Tag	Skills	Theme
v0.1.0-draft	16	Baseline
v0.2.0-draft	24	+ Web 3D handoff deep + Quality validators
v0.3.0-draft	32	+ Domain scene planners + Process discipline
v0.4.0-draft	40	+ Asset/sourcing + Color/format/output
v0.5.0-draft	48	+ Recipe/pack ops + Runtime path discipline

Assets (48 portable Skill ZIPs)

Every Skill is a stand-alone uploadable Anthropic Skill (≤200-char description, no version:/status:/tags: keys). Drag-drop straight into Claude.ai / Claude Desktop, drop into .claude/skills/ for Claude Code, or wire into any MCP-aware coding agent following per-target install docs in docs/install/.

Scope reminder

BlendOps still publishes one thing: the skills layer. Runtime paths (Path 1 Blender Lab MCP, Path 2 community ahujasid/blender-mcp, CLI fallback) remain upstream's deliverable — see docs/runtime-stack-strategy.md.

Verification

• npm run docs:check passes 342 active markdown files.
• npm run skills:export produces 48 ZIPs + 48 Claude Code folders deterministically.

v0.4.0-draft: 40 skills milestone

BlendOps v0.4.0-draft — 40 skills milestone

Roadmap progress checkpoint: 40 / 48 skills (83%) of the 3x baseline target.

What changed since v0.3.0-draft

Batch 5 — Asset / sourcing (4 skills)

• asset-license-checker — verify license compliance for every external asset before scene build/export
• asset-style-consistency-checker — flag style mismatch (PBR vs stylized, scale conflicts) before scene build
• asset-fallback-strategy — define plan-B sources, procedural fallbacks, and substitution rules
• asset-library-organization-planner — namespacing + folder + naming patterns for reusable libraries

Batch 6 — Color / format / output decisions (4 skills)

• color-management-decision — pick sRGB / Filmic / Standard / AgX / ACEScg per output context before render
• output-format-decision — choose PNG / JPG / WebP / OpenEXR / TIFF based on bit depth, alpha, compression
• resolution-aspect-decision — pick render resolution + aspect ratio per consumer use case; avoid render-then-crop
• cycles-vs-eevee-decision — Cycles / Eevee Next / Eevee legacy / Workbench based on quality, time, features

Each ships with SKILL.md + EVAL.md + 3 references/ files (24 new reference files).

Roadmap progress

Tag	Skills	Status
v0.1.0-draft	16	shipped
v0.2.0-draft	24	shipped
v0.3.0-draft	32	shipped
v0.4.0-draft	40	this release
v0.5.0-draft	48	next, final

Assets

40 portable Skill ZIPs (all Anthropic Skills frontmatter-compliant), one per skill.

Scope reminder

BlendOps publishes the skills layer only — runtime paths remain upstream's deliverable. See docs/runtime-stack-strategy.md.

BlendOps v0.3.0-draft - 32-skill milestone (Batches 1-4)

BlendOps v0.3.0-draft — 32-skill milestone (Batches 1-4 complete)

This draft milestone closes the first four batches on the path to the 48-skill (3x baseline) roadmap. Progress: 32 of 48 (67% of 3x roadmap).

Milestone	Skills	Status
v0.1.0-draft baseline	16	Shipped 2026-05-09
v0.2.0-draft (Batches 1+2)	24	Shipped 2026-05-09
v0.3.0-draft (this release)	32	Shipped 2026-05-09
v0.4.0-draft	40	Pending — Batches 5+6
v0.5.0-draft	48	Pending — Batches 7+8 (= 3x target)

What's new vs v0.2.0-draft

Batch 3 — Domain scene planners (4 skills)

Four narrowly-scoped scene planners that complement the existing product-hero-scene-planner:

• character-portrait-scene-planner — single-character portrait (subject, pose, expression, wardrobe, framing, lighting, background). References: portrait-framing-rules / portrait-lighting-patterns / style-consistency-rules.
• environment-establishing-shot-planner — wide outdoor / cityscape / landscape (locale, time, weather, atmosphere, FG-MG-BG depth layers, focal subject). References: time-and-weather-rules / wide-shot-camera-rules / depth-layering-rules.
• interior-architectural-scene-planner — interior arch viz (room dims, style, furniture, materials, natural+artificial lighting, camera). References: interior-architecture-rules / furniture-layout-rules / interior-camera-rules.
• product-grid-scene-planner — multi-product grid (e-commerce / catalog) with consistent lighting, camera, background, grid spacing. References: grid-layout-rules / grid-consistency-rules / grid-output-naming-rules.

Batch 4 — Process discipline (4 skills, Superpowers-inspired)

Four process / discipline skills that prevent scope drift, enforce gates, and pin "done":

• blender-checklist-driven-workflow — per-stage gates and acceptance checklists; work advances only when each previous gate Pass. Inspired by Anthropic Superpowers /test-driven-development analog. References: gate-template / stage-decomposition-patterns / failed-gate-handling.
• blender-stop-condition-decider — define one explicit "done" condition for a recipe before starting work, plus an out-of-scope list. Inspired by Superpowers /just-say-when-it's-done analog. References: stop-condition-templates / out-of-scope-patterns / done-vs-perfect.
• blender-scope-boundary-enforcer — detect mid-recipe scope creep ("can you also...") and force decline / agree-to-new-scope / defer; prevents silent absorption. References: scope-addition-classification / options-language-template / recurring-creep-patterns.
• blender-recipe-decomposer — break complex / over-scoped intent into 2-5 cohesive recipes, each with brief, stop condition, handoff. References: recipe-grouping-rules / dependency-patterns / recipe-naming-rules.

Release assets

The 32 ZIPs attached are produced by npm run skills:export from this exact tag. Each ZIP contains exactly one top-level skill directory with one SKILL.md conforming to Anthropic Skills frontmatter spec. Drag-drop them into Claude Desktop / Claude.ai / ChatGPT Skills UI to import.

Coding agents (Cursor, Claude Code, Cline, Continue, Zed, goose, Codex, OpenCode, Gemini, Antigravity, Copilot) use the README 30-second prompt to install project-locally — agent auto-detects mode.

Scope (unchanged)

BlendOps publishes the skills layer (32 skills + 4 laws + 1 pack + 3 bundles + 16 install docs + ZIP/package generation + CI guards). BlendOps redirects to upstream runtime paths (Blender Foundation Lab MCP, Anthropic Connector, ahujasid community, Blender CLI) — does NOT own them.

Path 1 + Path 2 user-reported verified by repo owner. CLI is decades-stable upstream. In-repo per-recipe runtime evidence files are optional operator-side provenance.

Verified at release time

• npm run docs:check validates 262 active markdown files.
• npm run skills:export produces 32 ZIPs + 32 Claude Code skill folders.
• All required headings + frontmatter spec present in every SKILL.md.

Coming next

• Batch 5 — Asset / sourcing (4 skills): asset-license-checker, asset-style-consistency-checker, asset-fallback-strategy, asset-library-organization-planner.
• Batch 6 — Color / format / output decisions (4 skills): color-management-decision, output-format-decision, resolution-aspect-decision, cycles-vs-eevee-decision.
• → v0.4.0-draft milestone (40 skills).
• Then Batches 7-8 → v0.5.0-draft final (48 = 3x target).

BlendOps v0.2.0-draft — 24-skill milestone (Batch 1 + 2)

BlendOps v0.2.0-draft — 24-skill milestone (Batch 1 + Batch 2 added)

This draft milestone closes the first two batches on the path to the 48-skill (3x baseline) roadmap goal. Status remains Draft v0; this release adds 8 new skills on top of the v0.1.0-draft baseline of 16.

Progress: 24 of 48 (50% of 3x roadmap)

Milestone	Skills	Status
v0.1.0-draft baseline	16	Shipped 2026-05-09
v0.2.0-draft (this release)	24	Shipped 2026-05-09
v0.3.0-draft	32	Pending — Batches 3 + 4
v0.4.0-draft	40	Pending — Batches 5 + 6
v0.5.0-draft	48	Pending — Batches 7 + 8 (= 3x target reached)

What's new vs v0.1.0-draft

Batch 1 — Web 3D handoff deep (4 skills)

The biggest unique value of BlendOps (web 3D handoff for non-Blender users) was previously concentrated in a single glb-web-handoff skill. Batch 1 splits it into 4 narrowly-scoped composable skills:

• glb-mobile-performance-budget — Set explicit triangle / texture / draw-call / file-size / animation budgets per device tier (low-end Android / mid-mobile / high-mobile / desktop fallback) before any export decision. References: budget-tiers, violation-rules, degradation-plan.
• glb-animation-handoff — Document GLB animation contract (skeletal vs morph, FPS, looping, root motion, blend transitions) with web-stack-specific caveats (Three.js / R3F / Babylon / model-viewer). References: animation-types, web-stack-caveats, transition-rules.
• three-fiber-component-shape-planner — Plan the React Three Fiber component shape (props, refs, Suspense, useGLTF cache, animation hooks, cleanup, fallback) before delivering a GLB. Spec only — no component code generated. References: component-shape-patterns, suspense-placement-rules, usegltf-caching-rules.
• preview-report-template-writer — Generate the structured preview-report Markdown that glb-web-handoff and non-blender-user-response-writer consume after a render or export run. References: report-schema, truth-label-rules, handoff-recipients.

Batch 2 — Quality validation deep (4 skills)

Four narrow domain validators that complement the existing broad blender-scene-quality-checker:

• material-quality-checker — Per-material PBR sanity (albedo / normal / roughness / metallic), texture budget, consumer-stack compatibility (allowed shader nodes, KHR extension confirmation). References: pbr-sanity-rules, texture-budget-rules, consumer-stack-rules.
• lighting-quality-checker — Per-light intensity ranges, shadow softness vs mood brief, color temperature consistency, HDRI / explicit-light balance patterns. References: intensity-ranges, shadow-softness-rules, hdri-balance-rules.
• composition-quality-checker — Per-aspect framing (rule of thirds, subject placement, headroom, lead room, negative space, hierarchy, lens / FOV plausibility). References: composition-rules, headroom-leadroom-rules, aspect-ratio-rules.
• polycount-budget-checker — Per-mesh + total-scene triangle counts vs pinned budget; worst-offender highlighting; advisory topology hints (no modeling performed). References: per-mesh-rules, topology-hints, instance-counting-rules.

All 4 quality checkers share a strict "validate, do not edit" boundary.

Release assets

The 24 ZIPs attached to this release are produced by npm run skills:export from this exact tag. Each ZIP contains exactly one top-level skill directory with one SKILL.md file conforming to Anthropic Skills frontmatter spec (≤200-char description, no version: / status: keys). Drag-drop them into Claude Desktop / Claude.ai / ChatGPT Skills UI to import.

For users in coding agents (Cursor, Claude Code, Cline, Continue, Zed, goose, Codex, OpenCode, Gemini, Antigravity, Copilot), use the README 30-second install prompt to install project-locally — the agent auto-detects mode (A = project-local, B = chat-only ZIPs, C = blocked-needs-input).

Scope (unchanged from v0.1.0-draft)

BlendOps publishes: 24 skills + 4 laws + 1 pack + 3 bundles + 16 install target docs + ZIP/package generation + CI guards.

BlendOps redirects to (does NOT own):

• Path 1 — Lab MCP (Blender Foundation) hosted from Anthropic Blender Connector or any other MCP client.
• Path 2 — Community ahujasid/blender-mcp (21K+ stars).
• CLI fallback (blender --background --python ..., decades-stable upstream).

Path 1 + Path 2 are user-reported verified by the repo owner. CLI is documented upstream as first-class. In-repo per-recipe runtime evidence files are optional operator-side provenance.

Verified at release time

• npm run docs:check validates 222 active markdown files.
• npm run skills:export produces 24 ZIPs + 24 Claude Code skill folders.
• All required headings + frontmatter spec present in every SKILL.md.

Coming next

• Batch 3 — Domain scene planners (4 skills): character-portrait, environment-establishing-shot, interior-architectural, product-grid.
• Batch 4 — Process discipline (4 skills): checklist-driven-workflow, stop-condition-decider, scope-boundary-enforcer, recipe-decomposer. → v0.3.0-draft milestone (32 skills).

BlendOps v0.1.0-draft — first skills layer release

BlendOps v0.1.0-draft — first Draft v0 skills layer release

This is the first tagged release of BlendOps. Status: Draft v0, not production, not stable. This release captures a complete skills layer; runtime execution is redirected to upstream paths and is not part of what BlendOps publishes.

What this release contains

Skills layer (16 skills, all at Expanded Draft v0)

Help / routing:

• blendops-help

Process / discipline (Superpowers + BMad inspired):

• blender-brainstorming
• intent-to-3d-brief-writer
• blender-asset-discovery-planner
• blender-troubleshooting
• pre-handoff-verification
• runtime-bridge-conflict-resolver

Runtime setup:

• official-runtime-setup-guide
• official-runtime-readiness-checker

Scene planning (domain):

• product-hero-scene-planner
• blender-composition-camera-planner
• blender-lighting-material-planner
• blender-scene-quality-checker

Evidence + handoff:

• render-export-evidence
• glb-web-handoff
• non-blender-user-response-writer

Laws (4): evidence-before-done, official-runtime-only, no-arbitrary-python-interface, non-blender-user-language.

Pack (1): product-hero-v0.

Bundles (3): skill-package, claude-desktop-manual, generic-project-local.

Install target docs (16): Claude Desktop, Claude Code, Cursor, Codex, OpenCode, Gemini, Antigravity, GitHub Copilot, Cline, Continue, Zed, goose, Ollama, LM Studio, Open WebUI, generic-project.

3-mode auto-detect install flow: Mode A (project-local), Mode B (multiple Skills ZIP for chat-only UIs), Mode C (blocked-needs-input). Documented in docs/ai-agent-quickstart.md + README 30-second prompt.

Scope clarification

BlendOps publishes the skills layer above. BlendOps redirects to the following upstream runtime paths but does not own them:

• Path 1 — Lab MCP (Blender Foundation, blender.org/lab/mcp-server) hosted from either Anthropic Blender Connector (Claude Desktop one-click toggle) or any other MCP client (Claude Code, Cursor, Codex, OpenCode, Cline, Continue, Zed, goose, etc.) configured manually. Requires Blender 5.1+.
• Path 2 — Community ahujasid/blender-mcp (21K+ stars). Requires Blender 3.0+.
• CLI fallback (blender --background --python ..., docs.blender.org/manual/.../command_line). Decades-stable upstream first-class CLI surface.

Path 1 + Path 2 are user-reported verified by the repo owner.

Release assets

The 16 ZIPs attached to this release are produced by npm run skills:export from this exact tag. Each ZIP contains exactly one top-level skill directory with one SKILL.md file, conforming to Anthropic Skills frontmatter spec (≤200-char description, no version: / status: keys). Drag-drop them into Claude Desktop / Claude.ai / ChatGPT Skills UI to import.

For users in coding agents (Cursor, Claude Code, Cline, etc.), use the README 30-second install prompt to install project-locally.

Truth labels

• Runtime status: Not Run (refers to in-repo per-recipe captured eval files; the upstream paths themselves are user-reported verified by repo owner).
• Artifacts: Not Produced (refers to in-repo captured render/GLB files; users produce real artifacts on their own machines via the redirect paths).
• Skill upload to chat-app UIs (Claude.ai, ChatGPT, Claude Desktop): manual user action; ZIPs are correctly shaped.

Verified at release time

• npm run docs:check validates 182 active markdown files against runtime model, frontmatter spec, evidence rules, and link integrity.
• npm run skills:export produces 16 ZIPs + 16 Claude Code skill folders from the same skill source.
• All 16 skills + EVALs registered in scripts/check-docs.mjs.

Not yet in this release (future deferred work)

• npx blendops installer (spec only at docs/install/installer-spec.md).
• Marketplace / plugin listings (gated on upstream listing acceptance).
• Web-ready 3D output implementation specifics (planning specs only at this point).
• Captured per-path runtime evidence files under docs/evals/ (optional operator-side provenance).
• Additional skills beyond the current 16 (planned roadmap toward ~48 skills total).

Inspirations

• Blender Lab MCP — official runtime path.
• ahujasid/blender-mcp — community runtime path.