ahujasid

Build a Blender scene from a text prompt with Claude Code Blocking out a scene in Blender by hand means hours of menus and modifiers. MCP for Blender lets Claude Code, Codex or another AI app drive Blender directly: it creates objects, places them, applies materials and checks its own work with renders. The short version is on the Deduped use-case page. Below is the route with the commands from the README. What you need Blender installed, an AI app that supports MCP, and macOS or Linux for the one-line installer. This walkthrough comes from the README, not a test run, so check the repo for newer instructions. One-command setup curl -LsSf https://www.mcp-for-blender.com/install.sh | sh If you already have uv, this does the same thing: uvx mcp-for-blender setup The setup asks which MCP clients to configure. It adds a blender entry to each and saves a .bak copy of every config it touches, so you can roll back. Manual route for Claude Code or Codex If the automatic setup misses your client, add the server yourself: claude mcp add blender uvx mcp-for-blender codex mcp add blender -- uvx mcp-for-blender uvx mcp-for-blender install-addon The last command installs and enables the Blender addon. Start the connection Fully quit and reopen your AI app, then open Blender. The addon starts its server, and the AI app can now see it. Ask for a scene Be concrete. A prompt shape that works: Using ahujasid/mcp-for-blender, build a low-poly campsite at dusk in the open Blender file: create the objects, apply materials, set up a camera and light, then show me renders from two angles and list what you created. Make it check itself Ask for multi-angle views or a render after each pass. Models misplace things in 3D often, and a render makes the mistake obvious. Then iterate in plain language: "move the tent left of the fire", "make the sky warmer". Safety note The server can run Python inside Blender, so save your file first, and use safe mode if you want to limit what it runs. Do not point it at a project you cannot restore. For another creative workflow with an agent, see the HyperFrames product video walkthrough. Want the setup commands for 50 AI repos people actually ship with? Get the free AI Repos That Actually Ship pack: dedupedai.com

Holy f*ck, Opus 5.5 just came for 3D designers' jobs... Opus 5.5 + Blender is a one-person animation studio that prints $$$. You give it one idea, and it builds a scene where every movement sets off the next one. The stack: 01 / Opus 5.5 + Blender MCP ahujasid/mcp-for-blender -> builds the scene and checks it 02 / Blender blender.org -> physics, keyframes and camera 03 / Remotion + FFmpeg remotion.dev ffmpeg.org -> edit, sound and final export The harness: brief -> event graph -> scene -> preview -> fix -> checkpoint -> export All event timings live in one file. After every approved change, save the editable scene. Hook it up to Claude Code: `claude mcp add blender uvx mcp-for-blender` `uvx mcp-for-blender install-addon` You need Blender + uv. Turn on the addon in Blender and start its server. The visual walks through this workflow on the reference clip. Save this, then read the full Opus 5.5 Motion Engineering guide ↓

@theozguratillay blender'in official mcp'sini mi kullaniyorsun yoksa bunu mu? ahujasid/mcp-for-blender

holy f*ck, Opus 5.5 just killed 3D designers jobs... Opus 5.5 + Blender = your own animation studio that prints $$$ one idea → a scene where every movement triggers the next the stack: 01 / Opus 5.5 + Blender MCP ahujasid/mcp-for-blender → build and inspect the scene 02 / Blender blender.org → physics, keyframes and camera 03 / Remotion + FFmpeg remotion.dev ffmpeg.org → edit, sound and export the harness: brief → event graph → scene → preview → fix → checkpoint → export keep the event timings in one file. save the editable scene after each approved change. connect Claude Code: `claude mcp add blender uvx mcp-for-blender` `uvx mcp-for-blender install-addon` requires Blender + uv. enable the addon in Blender and start its server the visual maps this workflow around the reference clip save this, then read the full guide for Opus 5.5 Motion Engineering ↓

then render your start and end frames from the set and feed them to seedance, kling or veo. both frames come from the same geometry. repo: ahujasid/blender-mcp what set would you build first?

Show HN: A 3D mesh playground for coding agents (three.js + manifold-3d) When GPT-6/Gemini launched recently, a lot of the 3D demos used blender-mcp (ahujasid/blender-mcp). It's cool, but you have to install blender locally and wire up plugins just to try a prompt or share a run. I built a browser-based 3D mesh playground so you can hook up your own coding agent over MCP (or run in-browser) directly in a tab. It runs on three.js + manifold-3d for blender-like boolean/mesh editing in JS, and gives the agent a render_preview tool to take multi-view screenshots (front/right/top/iso + custom camera angles + clay mode). That way the agent can look at what it built and fix the geometry, similar to how it iterates when building a website. The submitted link is a trace of Gemini reconstructing a Copenhagen canal scene from a photo: it hits a script error on the first pass, fixes the bug, and then goes through several rounds of checking 4-view and close-up render_preview screenshots to tweak the geometry.

Desk Racers Built From a Photograph, Purple Gradients Sent Back to the Kitchen, and a Security Backlog That Finally Got Help You’re reading The Agentic Rewind #2. --- A Noble Gas With a Long Attention Span Argon is a noble gas, which means it doesn't react to anything. And that, in a way, is rather wonderful. On October 13, Google released Gemini 4 Argon, and the idea is beautifully simple: give it a big task and walk away. Koray Kavukcuoglu (@koraykv) introduced it as a model trained to work longer and remember more than any Gemini before it. It plans, writes the code, checks its own work, and keeps going long after we'd have closed the laptop and gone to look at the stars. So what actually changed? Earlier Gemini models could only write 64K tokens before they had to stop. Argon shares one 1M-token window across everything it reads, thinks and writes. So a single call can return a whole multi-file change, not half a function. To go past 64K tokens, you run the request in background mode. let interaction = await client.interactions.create({ model: "gemini-4-argon", input: "Write a complete, rigorous proof " + "of the prime number theorem.", // past 64K output tokens needs background mode background: true, generation_config: { max_output_tokens: 262144, thinking_level: "high", }, }); Then you check back with client.interactions.get until it's finished, which is Google's polite way of saying go and make a coffee. What does a long run look like? Google's showpiece is Desk Racer. Argon took photos of a real desk and, working inside Antigravity, Google's agent-first coding tool, turned them into a playable miniature racing game. It built the 3D scene with Blender MCP, painted it with Nano Banana, and wired up the physics and engine sounds in Three.js. The track loops around the keyboard, the mouse and the desk lamp, a tiny universe built from a photograph. !Desk Racer still Every race starts with a countdown on the line, and you can try it yourself in AI Studio. !Desk Racer gameplay Does it actually ship better code? Google says its teams shipped over 35% more agent-written code with Argon, with nearly a third fewer rollbacks. It also says Argon rarely cheats now, like skipping a failing test, doing it just 2% of the time. Pricing starts at $2 per million input tokens and $10 for output, doubling on December 1. Our desks have never looked more productive. Mostly because they never had a racetrack before. This just made me want to play a round of Mario Kart (who are we kidding… more than one round, probably until I finish last three times in a row, hit by every evil banana and turtle shell going). 👉 Gemini 4 Argon --- !Gollum weekly usage limit meme Taste the Rainbow, Fail the Gradient Right, ask any coding agent for a landing page, and what do you get? A purple gradient, a centred headline, three feature cards, and a testimonial from someone who doesn't exist. It's bland, it's lazy, and we've all been served it a thousand times. Leon Lin (@LexnLin) shipped version 2 of Taste Skill on October 13, and its whole job is to send that page back before anyone sees it. It's an Agent Skill (a folder of instructions your coding agent loads when it needs them) for Claude Code, Codex and Cursor. What's new in version two? It's not the old recipe reheated. Instead of vague advice like "make it modern", it gives the agent a real library of designs, built from a study of 247 hand-picked sites. The agent is dealt three design directions and has to pick the one that fits this particular client. It also pulls real photos and maps from open collections like Wikimedia Commons and Openverse, with credits on the page. # install the v2 default skill npx skills add https://github.com/Leonxlnx/taste-skill \ --skill "design-taste-frontend" # deal three directions for the brief node <skill>/scripts/draw.mjs --sector <sector> \ --brief "<one line>" --sections <roles> --options 3 # the gate the agent has to pass node <skill>/scripts/lint.mjs <source files or folder> node <skill>/scripts/inspect.mjs <url or path/to/index.html> That last line is the head chef, tasting every plate before it goes out. inspect.mjs opens the page in a real browser on desktop, tablet and phone widths, and rejects it for gradients, shadows and other usual suspects. The agent looks at its own screenshots, fixes the worst three problems and tries again, up to three times. Does it actually look different? Here's a coffee shop page it made from a single brief, with a drawn storefront at the top. !Taste Skill coffee shop page And here's a techno night that prints its dates right onto an old photo of an industrial hall. !Taste Skill techno night page The catch is time: a build takes 30 to 45 minutes and a few hundred thousand tokens. It's also meant for public pages, not dashboards or app screens. Somewhere, a purple gradient is updating its LinkedIn profile. Shut it down. 👉 Taste Skill --- !Interstellar monorepo tokens meme 📦 Gemini 4 Argon Google's new frontier model is GA in the API, AI Studio, and Antigravity at an intro $2/$10 per million tokens until Nov 30… and it already untangled a ~6 million line folder of legacy Chrome code, so your monolith is next. 📦 HydraFusion Max Copilot's multi-model blender now pulls from Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, with parallel critics poking the actual workspace… Max cranks the reasoning for when "balanced" isn't the vibe. 📦 Codex 0.163 Mid-Turn Steering Send a follow-up while Codex is mid-response and it steers right away, plus searchable /config and read-only tools running in parallel… finally, a way to say "no, not that file" before it's too late. 📦 Cline Cloud Sessions Run a Cline task in a hosted sandbox on the repo and branch you pick, and it keeps going after you close your editor… rolling out gradually, and unpinned Cline-provider users now default to Gemini 4 Argon. 📦 Shop MCP Shopify's hosted MCP server lets any agent search 1B+ products, build a cart, and finish Shop Pay checkout without bouncing the shopper to a website… search doesn't even need an API key, which your credit card may find concerning. 📦 GitLab MCP Server GA Shipped with GitLab 19.5: semantic code search, pipeline debugging, toolsets, and Duo Agent Platform flows you can kick off from Claude Code or Cursor… all under your existing GitLab permissions, on every tier including Free. 📦 Copilot Agentic Autofix Now GA: assign a code scanning alert to Copilot and it explores the codebase, re-runs CodeQL to validate the fix, and opens a draft PR… up to 25 alerts at once, so the security backlog finally gets an intern. 📦 Copilot CLI Agent-Scoped Hooks 1.0.97 lets custom agents carry their own declarative hooks, and Autopilot now denies a hook's "ask" instead of freezing on a permission dialog nobody is watching… RIP the 3am prompt that patiently waited until 9am. 📦 Claude Code Retry Watchdog Cap v2.1.301 adds CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS, so unattended retry mode gives up on a usage limit that resets too far away… instead of sitting there all weekend like a dog at the door. 📝 How the Claude Code Team Does Code Review Claude writes about 90% of Claude Code, merged PRs more than doubled from June to September, and reverts fell over 40%… plus the tale of one PR that did 50+ review rounds overnight before anyone stopped it. !Goodbye GIF --- Authored by Luke Farrell and edited by Fran Ríos. If you’re enjoying this newsletter, why not help us grow? Share it with your friends, family, and even that one coworker who always “forgets” to update dependencies. Use this link: theagenticrewind.com.

Gemini 4 Argon Built a Tesla Roadster in Blender From 4 Pictures, Then Made It Fly Gemini 4 Argon (Google's model, also called Argon 4) built a white Tesla Roadster style car in Blender from 4 pictures and one prompt, then wrote a three.js launch show where the car climbs a ramp, hangs upside down and rolls back to land. The car cost about $1.54 (52 calls, 20.5 minutes) and the whole project about $6.96 (200 calls), both estimated from the measured token counts at Google's list price. This post has the 4-step setup, the exact prompt, every follow-up prompt with my typos, and the show itself, which you can play in your browser. {% embed youtu.be %} Play the launch show: kartit.net. Key facts (quick answer) | Fact | Gemini 4 Argon, Roadster in Blender | |---|---| | Date | 20 and 21 October 2026 | | Model | Gemini 4 Argon, medium effort (gemini-4-argon-medium) | | Agent | Google Antigravity command line (agy) | | Link to Blender | MCP for Blender, free add-on | | Input | 4 reference pictures (front, left, rear, right) + 1 prompt | | The car in Blender | 52 calls, 20.5 min, about $1.54 | | Car tokens | 2.06M input (1.79M cached), 81K output | | Everything, launch show included | 200 calls, about $6.96 | | All tokens | 16.3M input, 309K output | | Mesh | 7,420 polygons (measured in Blender) | | Speed | about 4 s a step on medium, 10 to 20 s on high | | Price basis | Google list price: $2/M input, $0.10/M cached, $10/M output incl. thinking | TL;DR - 4 pictures and 1 prompt in, a 3D car in Blender out, in 20.5 minutes. - The car cost about $1.54 for 52 calls (estimated at list price). - The car plus a flying three.js launch show and all its fixes: about $6.96 for 200 calls. - The finished mesh has 7,420 polygons. - Medium effort ran about 4 s a step; high effort ran 10 to 20 s. - The show took 12 prompts of direction; the car took 1 prompt and 1 fix. Play the launch show This is the actual page Argon wrote, and it runs in your browser. Drag to orbit the camera, or pick a shot from the bar at the bottom. On a phone, turn it sideways.  Play the launch show on kartit.net (three.js, runs on computers and phones). What did Gemini 4 Argon build? A 3D car in Blender that matches 4 reference pictures, then a cinematic web page around it. Blender is a free 3D app, the kind studios use for models and animation. Argon did not click around in Blender like a person. It drove Blender through MCP, a plug that lets the AI send commands to Blender and take screenshots of its own work. The agent was Antigravity, Google's AI coding agent, used here from the terminal (its command is agy). The car has a white wedge body, a black lower body, a dark glass canopy, a thin white light bar at the front, a red light bar at the back and black 6-spoke wheels. The finished mesh has 7,420 polygons. What were the 4 pictures? Four flat views of the same car: front, left side, rear, right side. That is all Argon got to look at. !The 4 reference pictures given to Gemini 4 Argon: Tesla Roadster style car from the front, left, rear and right How close is the 3D car to the pictures? Close on shape and proportions, simpler on detail. Each pair below is the reference picture on the left and Argon's Blender model from the same angle on the right. !Front view: reference picture next to Gemini 4 Argon's Blender model with the white front light bar !Left side: reference picture next to Gemini 4 Argon's Blender Roadster model with 6-spoke wheels !Rear view: reference picture next to Gemini 4 Argon's Blender model with the red rear light bar and diffuser !Right side: reference picture next to Gemini 4 Argon's Blender Roadster model The body, the canopy, the light bars and the 6-spoke wheels all landed where the pictures put them. The surfaces are flatter than in the pictures and the side panels are plainer. It is not as polished as a car made by a 3D artist, but it is good enough for a game or a short film. !Gemini 4 Argon's Roadster model in Blender, wireframe view showing the 7,420 polygon mesh !Gemini 4 Argon's Roadster model in Blender, solid shaded rear three-quarter view How do I connect Gemini 4 to Blender? Four steps, all free tools. 1. uvx mcp-for-blender setup and type 7 to pick Antigravity only. This installs MCP for Blender, the free add-on that connects the AI to Blender. 2. Open Blender. The add-on starts its server by itself. To see it, press N and open the tab "MCP for Blender". 3. agy mcp list should show: blender enabled. 4. agy --model gemini-4-argon-medium starts the agent on Gemini 4 Argon at medium effort. What was the prompt? One prompt, word for word: > Build this roadster in Blender with your Blender tools. Look at the 4 reference views first: refs/front.png, refs/left.png, refs/rear.png, refs/right.png. Start from an empty scene. White wedge body, black lower body, dark glass canopy, a thin white front light bar, a red rear light bar, black 6-spoke wheels. Check your work with viewport screenshots from the same 4 angles and fix what does not match. Finish with a nice 3/4 front view in the viewport. The important line is the one that asks Argon to check its own screenshots from the same 4 angles. That is what made it compare its model with the pictures and fix the differences without me. What did I ask after that? Short follow-ups, typed fast. Here they are in order, typos included: 1. "there is some black shapes on windshield smoothen design" 2. "open this in a three js and make a cinematic rpesentation of the car" 3. "the car should have thursters that throw vapor like in bottom to elevate the car animation" 4. "too much white glow" 5. "car supposed to hve doors open like this" (with a picture) 6. The stunt, in 12 prompts. I also gave it the link to the Roadster launch video. Some of them: - "you need to replicate this full animation scene" - "review the car animation it is supposed to speed then rotate and land from the jump" - "it has sucktion mecanism to stay" - "mimic drone style filiming car when it is going up then jump rotate and land" - "the car supposed to stay inverted till it jumps then it has right and left thrusters to rotate it back to land" - "fix the animation is wrong" Three.js is a free code library that draws 3D scenes inside a web page. Argon wrote the whole show around its Blender car: the C-ramp, the suction fan that holds the car upside down, the roll thrusters, the drone camera and the landing. One last fix round on the evening of 21 October lit the fourth act better, made the climb, hang and roll easier to read, eased the landing with a camera shake and put the doors on proper hinges. !Gemini 4 Argon's three.js launch show: the Roadster hanging upside down under the C-ramp, with downforce and altitude gauges How much did it cost? About $1.54 for the car and about $6.96 for everything. Both are estimates: I measured the tokens of every Gemini call and priced them at Google's list price. It is not a bill. | Part | Calls | Input tokens | Output tokens (incl. thinking) | Estimated cost | |---|---|---|---|---| | The car in Blender (20.5 min) | 52 | 2.06M (1.79M cached) | 81K | $1.54 | | Everything: car, launch show and all fixes | 200 | 16.3M | 309K | $6.96 | Prices used for Gemini 4 Argon: $2 per million input tokens, $0.10 per million cached input tokens, $10 per million output tokens, thinking included. Most of the car's input was cached: 1.79M of the 2.06M tokens. Cached input is the part of the conversation the model has already seen, and it is billed at $0.10 per million instead of $2. That is why 2 million tokens of reading cost so little. Medium or high effort? Medium. In this recording medium effort took about 4 seconds a step. High effort took 10 to 20 seconds a step. For a build where the agent takes a screenshot, looks, and changes one thing at a time, that difference adds up fast. Caveats - This is one build on one day, not a benchmark. Another car or another prompt will cost more or less. - The costs are estimates from measured token counts at list price, not an invoice. - The car is a Roadster style model built from 4 reference pictures, not an official Tesla model. - The launch show needed 12 prompts of direction. The car needed 1 prompt and 1 fix; the show is where the work was. My take What surprised me was the self-check. I asked Argon to take screenshots from the same 4 angles as my pictures and fix what did not match, and it did exactly that, round after round, while I watched. A 3D car for about $1.54 is cheap enough to try an idea before paying a 3D artist. The show was the opposite: the animation needed a director. Short, badly typed prompts were enough, as long as I said what was wrong. Sources - Measured token counts for every Gemini call, session of 20 and 21 October 2026, plus the fix prompt of 21 October 2026, priced at Google's Gemini 4 Argon list price (ai.google.dev/gemini-api/docs/pricing, read 21 October 2026). - Polygon count measured in Blender on the final model. - MCP for Blender (free add-on): ahujasid/mcp-for-blender - Playable launch show: kartit.net --- Originally published on kartit.net. Independent project by Anass Kartit. Not affiliated with Google or Tesla. New builds every week on YouTube: subscribe. Your turn: what should Gemini build in Blender next? Tell me in the comments and I'll try it.

日本語の指示だけでBlenderに3Dシーンを作らせてみた(mcp-for-blender) 本日のGitHub Trending(Python/全言語)で+2,098star/日という伸びを見せていたahujasid/mcp-for-blender(MIT License、旧称blender-mcp)を実際に手元の環境で動かしてみた。 AIクライアント(Claude Desktop/Claude Code等)からModel Context Protocol(MCP)経由でBlenderのPython API(bpy)を直接実行させ、シーン作成・アセット取り込み・3Dモデル生成を自然言語で行えるというツールである。筆者はBlenderのGUI操作にあまり詳しくないが、「日本語で指示するだけでBlenderに何かを作らせられる」という体験がどこまで本当なのか、実際に検証してみた。 前提: この記事で検証した範囲 Poly Haven等の無料アセット連携は対象外(時間の都合でシーン生成のコア動作確認を優先した) Sketchfab/Hyper3D/Meshy/Tripo等、APIキーが必要な有料3D生成連携は最初から対象外 GPUは未使用(Blender 4.0.2のEEVEEレンダラーをCPUで実行) 追加の有料APIキーなしで、Blender 3.0+・Python 3.10+・uvさえあれば検証できる範囲に絞っている。 まずBlenderが無い環境から始まった 検証環境にBlenderが入っていなかったので、まずインストールした。 $ apt-get install -y blender $ blender --background --version Blender 4.0.2 mcp-for-blenderのアドオン(addon.py)を有効化しようとすると、requestsモジュールが見つからずに失敗した。 ModuleNotFoundError: No module named 'requests' 配布パッケージ版のBlenderはrequestsを同梱していないことがある。Blenderが使うPythonのパスに向けて明示的にインストールし直すことで解決した。 $ python3.12 -m pip install --target ~/.config/blender/4.0/scripts/addons/modules requests ハマりポイント: blender --backgroundでは動かない 次にblender --background(ヘッドレスモード)でアドオンを有効化してみたが、ソケットサーバー自体が起動しなかった。アドオンのソースコードを読むと、これは意図された仕様だとわかる。 def start(self): if bpy.app.background: print("BlenderMCP: cannot start server in background mode (blender -b) - commands would never execute\n" "BlenderMCP: run Blender with a GUI, or use a virtual display: xvfb-run -a blender") return MCPサーバー側のタイムアウトメッセージにも同じ助言が書かれている。 If Blender is running headless (blender -b), commands never execute; run Blender with a GUI or via 'xvfb-run -a blender' instead つまりツール自身が「GUIなしサーバーで動かしたいなら仮想ディスプレイ(Xvfb)を使え」と明示的に案内している。これに従い、xvfb-run -a blenderでBlenderを(GUIモードのまま)仮想ディスプレイ上で起動したところ、アドオンのソケットサーバーが正常に立ち上がった。 $ nohup xvfb-run -a blender > blender_gui.log 2>&1 & $ cat blender_gui.log BlenderMCP addon registered Server thread started BlenderMCP server started on localhost:9876 生のTCPソケットで疎通確認すると、デフォルトシーン(Cube/Light/Camera)の情報が正しく返ってきた。 $ python3 -c " import socket, json s = socket.socket(socket.AF_INET, socket.SOCK_STREAM) s.connect(('localhost', 9876)) s.sendall(json.dumps({'type': 'get_scene_info', 'params': {}}).encode('utf-8')) print(s.recv(8192).decode()) " {"status": "success", "result": {"name": "Scene", "object_count": 3, "objects": [{"name": "Cube", ...}, {"name": "Light", ...}, {"name": "Camera", ...}]}} 実際にMCPクライアントとして繋いでみる この検証環境にはClaude Desktop等のGUIクライアントが無いため、Pythonのmcpパッケージ(stdio_client/ClientSession)で簡易MCPクライアントを自作し、uv run mcp-for-blenderをstdio経由のMCPサーバーとして起動して実際にMCPプロトコルで接続した。 from mcp import ClientSession, StdioServerParameters from mcp.client.stdio import stdio_client server_params = StdioServerParameters(command="uv", args=["run", "mcp-for-blender"], cwd="mcp-for-blender") async with stdio_client(server_params) as (read, write): async with ClientSession(read, write) as session: await session.initialize() tools = await session.list_tools() 公開されているツール一覧はこの通りだった。 ['disable_telemetry', 'execute_blender_code', 'generate_3d', 'get_addon_status', 'get_scene_info', 'import_asset', 'look', 'record_trajectory_feedback', 'search_assets'] 日本語の指示でシーンを作らせる 以下の日本語指示をuser_promptとして渡し、execute_blender_codeツールでBlender Pythonコード(壁4枚・床・松明・カメラを配置するコード)を実行させた。 低ポリゴンのダンジョンシーンを作って。石の壁で囲まれた部屋と、中央に松明を1本置いて、カメラで部屋全体が見えるようにレンダリングして。 実行前のget_scene_infoはデフォルトシーンのまま。 Scene 'Scene' | 1 camera, 1 light, 1 mesh | active Cube | selected Cube | mode OBJECT Cube | mesh | at (0.0, 0.0, 0.0) | size 2.0x2.0x2.0 execute_blender_codeを実行すると、こう返ってきた。 Code executed successfully: SCENE_BUILD_RESULT:{'object_count': 8, 'object_names': ['Floor', 'RoomCamera', 'Torch_Flame_Light', 'Torch_Handle', 'Wall_East', 'Wall_North', 'Wall_South', 'Wall_West']} 実行後のget_scene_infoでは、指示通り壁4枚・床・松明(柄+暖色の点光源)・カメラが揃っていることをMCP経由のテキスト応答だけで確認できた。 Scene 'Scene' | 1 camera, 1 light, 6 mesh | active RoomCamera | selected RoomCamera | mode OBJECT Wall_North | mesh | at (0.0, 3.0, 1.5) | size 6.0x0.4x3.0 Wall_South | mesh | at (0.0, -3.0, 1.5) | size 6.0x0.4x3.0 Wall_East | mesh | at (3.0, 0.0, 1.5) | size 0.4x6.0x3.0 Wall_West | mesh | at (-3.0, 0.0, 1.5) | size 0.4x6.0x3.0 Floor | mesh | at (0.0, 0.0, 0.0) | size 6.0x6.0x0.0 Before/Afterのレンダリング画像 execute_blender_code内でbpy.ops.render.render(write_still=True)を呼べば、MCP経由でBlenderのレンダリング結果を画像ファイルとして保存できる。デフォルトシーンと、日本語指示で生成した後のシーンをそれぞれレンダリングした。 Before(デフォルトのCubeシーン): After(日本語指示で生成した石壁の部屋+松明): 余談だが、カメラの初期配置(部屋の外から水平に狙う位置)では、厚み0.4のソリッドな壁がカメラと室内の間を完全に塞いでしまい、無地のグレー画像になってしまった。これは筆者のシーン構築コード側のカメラ配置ミスであり、ツール自体の不具合ではない。天井の無い部屋を上空から見下ろすアングルに置き直したところ、上記のAfter画像のように室内と松明の光が映るようになった。同じソケット接続上で何度でもコードを再実行・再レンダリングできる手軽さも確認できた。 追加検証: ローカルPCの実際のAIクライアントから接続してみた ここまではクラウドの実行環境にGUIクライアントが無く、MCPクライアントの役割をPythonの自作スクリプトで代替していた。後日、人間のローカルPC(Windows 11 + Blender 4.5)で、Claude Codeセッション自身を実際のMCPクライアントとして接続する検証を追加で行った。 公式のWindowsインストーラ手順はシンプルだった。 powershell -ExecutionPolicy ByPass -c "irm astral.sh | iex" uvx mcp-for-blender setup setupコマンド1つで、Blenderアドオンの導入・有効化と、見つかったMCPクライアント(Claude Desktop・Claude Code・VS Code・Antigravity)への設定登録がまとめて完了した。Blenderを通常のGUIモードで起動すると、追加の手動操作なしにport 9876でソケットサーバーが自動起動することも確認できた。これは、クラウド環境での検証(2節)で見た「blender --backgroundでは動かない」という仕様と一致する挙動だった。 $ claude mcp list blender: C:\Users\matsu.local\bin\uvx.EXE mcp-for-blender - ✔ Connected 実際にClaude CodeからMCP接続した状態でexecute_blender_codeを呼び、トーラスと暖色のポイントライトを追加してレンダリングしたところ、クラウド環境と同じようにシーンの変化(1 mesh/1 light → 2 mesh/2 light)と画像出力を確認できた。クラウドのLinux+Xvfb環境と、ローカルのWindows+通常GUI環境という違いはあったが、ツールの挙動そのものは一致していた。 まとめ mcp-for-blenderの「MCPクライアント→MCPサーバー→Blenderアドオン→bpy実行」という全スタックは、GPUや有料APIキーなしのLinuxサーバー環境でも、Xvfb経由であれば実際に動作した。 blender --backgroundでは動かないが、これはツール側が意図的に拒否する仕様であり、エラーメッセージ・ソースコード双方がxvfb-run -a blenderを公式な代替手段として明記していた。ツールの案内に従うだけで解決できる種類の制約である。 日本語の自然文の指示から、MCP経由でBlenderのシーン(オブジェクト・マテリアル・ライト・カメラ)を実際に変更でき、レンダリング画像として結果を確認できた。 追加検証として、ローカルPCの実際のAIクライアント(Claude Code)からも同じスタックが問題なく動作することを確認した。クラウド環境での代替検証と、ローカルでの実クライアント検証の両方で、結論は変わらなかった。 Poly Havenのアセット取り込みについては、今後別の機会に試してみたい。 出典 リポジトリ: ahujasid/mcp-for-blender (MIT License) PyPI: pypi.org

MCP for Blender: Connect Blender to any LLM (31k stars) What it does All text below is quoted directly from the project README. >Community plugin to control Blender 3D with any LLM of your choice. Not affiliated with the official Blender Foundation. Features: - Two-way communication: Connect Claude AI to Blender through a socket-based server - Object manipulation: Create, modify, and delete 3D objects in Blender - Material control: Apply and modify materials and colors - Visual verification: Multi-angle, camera, wireframe, X-ray and animation-strip views, plus renders, so the AI checks its own work - Code execution: The AI writes Blender Python directly - Asset & model generation: Poly Haven assets, Sketchfab models, Poly Pizza low-poly models, and AI-generated 3D models via Tripo, Hyper3D Rodin and Hunyuan3D >One command sets up your AI apps and the Blender addon: curl -LsSf https://www.mcp-for-blender.com/install.sh | sh Repository ahujasid/mcp-for-blender Repository stats |Metric|Value| |:-|:-| |Stars|31,032| |Forks|2,821| |Last release|None| |Languages|Python 98.1%, HTML 1.4%| |Contributors|47| |AI contributor accounts (Claude / Codex / Cursor / Copilot)|cursoragent: 3, claude: 7| |AI co-authored commits|Claude: 7, Cursor: 3| Main contributor |Field|Value| |:-|:-| |GitHub|@ahujasid| |Commits to this repo|152 (top contributor)| |Bio|Building open-source AI integrations for creative software| |Website|whimsy.is| |X / Twitter|@sidahuj| |Followers|778| |Public repos|16| |On GitHub since|2015-04-05| ------------------------------------------------------------------------------------------ Discover the best GitHub repos without the endless scrolling. Get top repos organized by category (AI, DevOps, Databases & more), plus key takeaways from developers' most useful comments - all in one free, twice-weekly newsletter. 👉 Get the best repos & community insights - Free