hitkeeen/blender-ai-mcp-addon is a trending Blender extension that runs an in-editor MCP server so Claude, GPT, Gemini, Ollama, or OpenCode can read the scene, call operators, and author bpy Python — with visual chat in the viewport. GitHub’s November 2026 trending list showed about 2,181 stars on a repo created November 12, 2026, signaling demand for 3D copilots beyond flat IDE assistants.
The addon sits next to Qwen 3D coders and VibePrevisBench in the agent-meets-3D lane, but it targets working bpy pipelines instead of one-shot mesh exports.
TL;DR
| Feature (per repo description) | Why it matters |
|---|---|
| MCP server inside Blender | Reuse agent stacks from Claude Code or Codex |
| Visual chat | Grounds model answers in viewport context |
| Multi-agent workflows | Split modeling, rigging, scripting across roles |
| Local + cloud models | Ollama path for air-gapped studios |
Architecture in plain language
MCP hub cabling a chart, a gear and agent nodes for creative tooling
Blender already embeds Python via bpy. The addon adds:
- MCP transport — external agents discover tools (scene queries, operator invocations, script execution).
- Visual chat UI — artists see prompts/responses beside the 3D view, reducing copy-paste into terminals.
- Multi-agent hooks — README markets coordinated workflows (for example layout agent plus scripting agent).
That mirrors how FFmpeg MCP skills wrap CLI utilities, but bpy can mutate scene data irreversibly — backups first.
Install and connect (high level)
Exact steps live in the repo README (eight commits at fetch time — fast-moving). Typical pattern:
- Download or clone github.com/hitkeeen/blender-ai-mcp-addon.
- Install the addon through Blender Preferences → Add-ons per README path.
- Enable the MCP server panel and note host/port or stdio settings.
- Point your MCP client (Claude Desktop, Cursor, OpenCode, etc.) at the server definition — same JSON patterns as Google Stitch skills.
- Start with read-only tools (list objects, inspect materials) before enabling execute bpy.
Test on a duplicated .blend file. bpy can batch-delete objects or rewrite drivers in one turn.
Use cases builders actually want
| Job | How MCP helps |
|---|---|
| Blocking + kitbash | Agent places arrays, you tweak proportions |
| Material variants | Parameter sweeps via scripted nodes |
| Pipeline glue | Export glTF/USD steps tied to Qwen GLB workflows |
| Rigging helpers | Generate bone naming conventions and constraints drafts |
Avoid treating output as shipping rigging without artist sign-off — benchmark coverage in embodied AI gaming shows agents still fail long-horizon spatial tasks.
What artists ask before enabling bpy execution
Will the addon delete my scene? Any bpy script can. Enable auto-save and use File → Recover → Auto Save.
Which Blender version is supported? Check the repo Issues and Releases — eight commits at first fetch meant rapid iteration; pin a tag for teams.
Can I use local Ollama only? Description explicitly lists Ollama among supported clients — useful for NDA assets.
How is this different from BlenderGPT forks? MCP standardizes tool schemas so the same agent drives Blender, filesystem, and ticketing — compare A2A vs MCP.
Does it replace add-ons like CAD Sketcher? No — it orchestrates operators; precision modeling still needs human judgment.
Multi-agent workflow pattern (example)
A bell pulse triggering coordinated gears as a multi-agent workflow metaphor
A practical studio split:
- Layout agent — places cameras and reference empties via MCP scene queries.
- Shading agent — adjusts node trees within material slots the lead artist locked.
- Export agent — writes glTF/USD with naming conventions for the game repo.
Each role gets different MCP tool allowlists — mirroring AWS manufacturing agent taxonomy on a creative pipeline.
Security and licensing
Community addons are unsigned supply chain. Before enterprise rollout:
- Diff the addon against tagged releases; pin commits.
- Ban network imports inside agent-generated bpy unless reviewed.
- Separate NDA assets from cloud model calls — same lesson as NASiko token gateways.
Read the repo LICENSE (present in tree) for redistribution rules on commercial work.
Tool surface (conceptual)
Without copying the entire README, expect MCP tools grouped roughly as:
| Tool family | Typical action |
|---|---|
| Scene inspection | List objects, materials, modifiers |
| Viewport / render | Capture previews for vision models |
| Operator invoke | Call Blender builtins with parameters |
| bpy scripts | Execute Python the human approved |
Exact names change between commits — introspect via MCP list_tools after install.
Prompt patterns that work
- “List modifiers on the active object” — low risk sanity check.
- “Duplicate lights to three-point setup, do not delete mesh data.”
- “Export glTF with these naming rules:” — pair with Qwen GLB coder.
Avoid “optimize my scene” without constraints — models may merge objects you needed separated.
Studio rollout checklist
- Pin addon git SHA in
requirements-blender.txtequivalent. - Run render farm jobs without MCP enabled (batch mode).
- Keep NDA assets on Ollama or air-gapped endpoints.
- Log MCP transcripts for client billing like Together agent evals culture expects.
bpy execution sandbox ideas
Studios uncomfortable with raw bpy can:
- Wrap MCP
execute_scriptbehind a human approval modal in a forked addon. - Whitelist operator names instead of arbitrary Python.
- Render on a duplicate blend file synced hourly from production.
None of that ships in the upstream repo — it is organizational policy you layer on top.
Benchmarks and evaluation
Compare agent output against VibePrevisBench discipline: measure task success, not vibes. A model that writes plausible bpy but deletes modifiers fails real pipelines.
Licensing and AGPL conversations
If your game ships GPL/AGPL tooling, verify Blender’s GPL interaction with your proprietary plugins before agents merge code into shared repos. Legal review beats “the model said it was fine.”
Classroom use
Instructors teaching 3D + AI can demo MCP with open assets only, mirroring Caltech homework grading debates about what students may automate. Require students to explain every bpy diff the addon logs.
Rendering farm integration
Large studios may run batch renders on a farm while MCP assists layout artists on workstations. Never point agents at production farm controllers without network segmentation — a bpy script could enqueue million-sample test renders that burn farm budgets.
Material libraries and asset linking
Agents that rename objects can break relative texture paths. Enforce pack all into blend before agent sessions, or use Blender’s link/override patterns documented in official manuals. Our Qwen GLB export post shows how naming conventions propagate to game engines — keep agents inside those conventions via prompt templates.
Vision-model caveats
When the addon feeds viewport captures to multimodal models, lighting and color management change what the model “sees.” Artists should freeze view transform during agent reviews or risk wrong material assignments.
Community support expectations
Eight early commits mean Discord/GitHub Issues are the support channel — not SLAs. Enterprise studios should fork and staff maintenance or wait for stable v1.0.0 tagging.
Trending velocity context
About 2,181 stars in four days signals interest, not maturity. Eight commits at first fetch is an early project — pin versions before semester-long courses depend on it.
Pipeline handoff to game engines
After MCP-assisted modeling, teams export to Unity, Unreal, or Godot using each engine’s import settings. Document scale factors and forward axis in a README agents can read — reduces “giant mesh” surprises when MiHoYo VLA gaming research teams test agents in-engine.
Audio and VFX scope boundary
This addon targets 3D scene automation, not sequencer video editing. For ffmpeg agent skills, use our FFmpeg MCP trending repo guide instead of forcing Blender bpy into video pipelines.
Long-term maintenance bet
If Blender 4.x API changes break operators, community addons need rapid fixes. Budget maintainer time or contribute PRs upstream — starring a repo does not ship production support.
Node-based materials and geometry nodes
Blender’s Geometry Nodes graphs are harder for LLMs than plain mesh operators. Start agents on mesh edit mode tasks before asking them to rewrite node trees — errors there are silent until render time. When you graduate to nodes, snapshot Group inputs names agents are allowed to touch.
Collaboration with non-technical stakeholders
Visual chat lowers the barrier for directors to request camera moves without learning bpy. Keep a human approver on destructive operators so art directors do not accidentally approve mesh deletes they thought were previews.
Version control with binary blends
Git LFS stores blends but diffs are opaque. Pair MCP sessions with Descriptive commit messages listing which agent ran which tool — future you will need audit trails when a client asks why a rig broke on Tuesday.
Closing recommendation
Studios evaluating Muse or Tripo style generators should still keep bpy agents for pipeline glue tasks generators skip, such as renaming collections or fixing scale on CAD imports.
Treat hitkeeen/blender-ai-mcp-addon as a prototype copilot layer: powerful for iteration, insufficient alone for shipping without artist review. Pair it with your studio’s existing pipeline validators and official Blender documentation for operator names agents should never call unsupervised.
Accurate as of November 17, 2026. Feature descriptions follow the hitkeeen/blender-ai-mcp-addon GitHub README; verify MCP tool names on your pinned commit before documenting internal runbooks for artists.
