Vol. 1 · No. 338Blender 3.6 / 4.2 / 4.5 / 5.0 / 5.2 / 5.3Cycles · EEVEE
JANCTION
Render · Blender on RTX PRO 6000 · Tokyo
0.1 JPYper GPU-secondFirst 100 JPYfree, no card
Front page / Agent-generated 3D: from a text prompt to a rendered image or MP4

Agent-generated 3D: from a text prompt to a rendered image or MP4

An agent does not need Blender to make 3D: it writes a bpy script that builds the scene, sends it to JANCTION Render, and gets a 720p preview back in a few GPU seconds. It looks at the image, fixes the script (camera, scale, lights, materials), repeats until it looks right, asks the user, then renders the final frames or an MP4. The preview is cheap on purpose, so the loop costs almost nothing. AI agents connect at https://render.janction.jp/mcp (remote MCP with OAuth; connecting creates a free key).

Updated 2026-11-17

See also: Render an AI-generated 3D model (Meshy, Tripo, Rodin, Hunyuan3D) into product shots and a turntable

What it does Your own .blend, bpy script or 3D file (glTF/GLB, FBX, USD, OBJ...) rendered to images or video on a cloud NVIDIA GPU, from an AI app, the terminal or the browser. The first preview needs no payment.
What it does not do It does not speed up the Blender viewport on your machine, does not take non-Blender video, simulations or general GPU jobs, and an AI that is not connected cannot render on its own.

Paste your AI's script and tryTry your own file for free

Browser only; no connector or CLI needed. A 100 JPY welcome credit for a new key (14 days, no card; one per network every 30 days, further keys on the same network share what is left), and 400 JPY more for saving a card that is not charged (once per card); after that, previews stay free up to 2 GPU-minutes a day and finals cost 0.1 JPY per GPU-second.

Measured, not estimated

One real loop from a prompt to a finished video, each step measured on the production GPU (NVIDIA RTX PRO 6000 Blackwell):

machine / routesettingsGPU secondscostround trip
an agent wrote the bpy script for "a matte ceramic mug on a wooden table, soft studio light, slow turntable" (HTTP API, 2026-10-12)preview 1: frames 1, 8, 16, 24 tiled, studio light4.26 GPU-swithin the free previews8.1 s
the preview's verdict: check — "cut by the frame edge: Cylinder", with bpy code for a camera position that frames every object; the agent appended it and previewed againpreview 24.21 GPU-swithin the free previews5.9 s (verdict: ok)
the final, 24 frames as MP4 (three chunks of 8 frames rendered in parallel)1280x720 / 6441.1 GPU-sabout 4 JPY (covered by the welcome credit here)15.3 s
a longer final from Claude Code (stdio MCP)48 frames, 1920x1080 / 128322 GPU-sabout 32 JPY335 s

The whole loop took 49.6 GPU-seconds, about 5 JPY at the per-second price, and a new key's welcome credit covered it. The same loop with its exact numbers is on /benchmarks.

Sources: /benchmarks (2026-09-30 to 2026-10-01), examples/one-command (2026-10-10), a laptop measured on 2026-10-11. Individual runs, not a guarantee; 0.1 JPY per GPU-second after the welcome credit.

Rendered on the production GPU

A 5-second animation an AI wrote from one sentence: a paper boat at dawn (120 frames, EEVEE, MP4)
138.2 GPU-s, 13.8 JPY, round trip 39 s, 2026-10-20
render_final(scene_url="https://render.janction.jp/samples/ex23_paper_boat.py", frame_start=1, frame_end=120, width=1280, height=720, samples=64, engine="eevee", output="mp4", fps=24)

The script, settings and GPU seconds are on each example's page. Prices at 0.1 JPY per GPU-second; 0 while the free credit covers it.

Inspect the input and output, then reproduce

Install uv. Give the linked bpy sample to your agent, request a preview and an estimate, then approve the final render. The command replays the saved sample; agent scene generation is not included in these timings.

Measured2026-10-10
Input and licenseMIT
Settings48 frames, 1280x720, EEVEE, MP4
GPU seconds54.1
Round tripnot recorded
At the current per-second rate5.41 JPY
OutputPNG / MP4 preview
Source record and procedurehttps://github.com/JasmyLab-JANCTION/janction-render/tree/main/examples/one-command
uvx janction-render render https://render.janction.jp/samples/ex12_logo_ring_spin.py --frames 1-48 --size 1280x720 --engine eevee --output mp4 --wait

Individual historical observations. Queueing, transfer, input and software versions can change elapsed time. Prices are illustrative, before free credit, whole-yen rounding and outcome ceilings; check /v1/estimate and /v1/pricing for your request.

The loop

  1. scene_info reads the scene without rendering: cameras, frame range, objects, missing files.
  2. render_preview renders 1-4 frames at 720p in a few GPU seconds and returns the image, so the agent can look, fix the script and try again.
  3. render_estimate says how long the final render takes ("about 3 minutes") and whether it fits the free GPU time left (the welcome credit while it lasts, then up to 2 preview GPU-minutes a day).
  4. render_final renders the frames on GPUs and joins them into an MP4; render_status reports the remaining time; render_download returns the files or links.

In practice an agent needs two or three previews before a scene looks right. Each is a few GPU seconds, so a whole session fits easily in the welcome credit, and after it in the free previews (up to 2 GPU-minutes a day).

What agents get wrong, and what to tell them

  • No camera or light. The preview is black or shows nothing. Add a camera and set scene.camera; pass environment='studio' for HDRI lighting. The service also adds an automatic camera and sun and says so in warnings.
  • Wrong scale or framing. Objects at 0.01 m or 100 m look like dots or walls. Keep the scene around 1-10 m and let orbit=true frame it automatically for a turntable.
  • Blender 4.x habits. mat.use_nodes = True is an error in 5.0 (always on); action.fcurves is gone; use keyframe_insert. The bpy rules list the rest.
  • Calling the render itself. Scripts must not call bpy.ops.render.render or change render settings; the service sets engine, resolution and samples from the request.
  • Images from the internet. The container has no network. Textures go through asset_urls (or polyhaven:<id> after asset_search), then os.environ['JR_ASSETS_DIR'] in the script.

Prompt patterns that work

Using JANCTION Render, build a glossy red cube rotating on a gray floor and show me a 4-frame preview.
If it looks right, render 24 frames at 1280x720 as an MP4.

To render a .blend or 3D file that is on your computer, drop it at /upload to get a 12-hour link (no sign-up) and tell the AI "Render this file: <link>". To test the connection first: "Render a preview of https://render.janction.jp/samples/cube_scene.py".

"Make a product shot of a ceramic mug on a wooden table, studio lighting, preview it, then give me a 48-frame turntable MP4." · "Build a low-poly city block with 20 buildings and a camera fly-through; preview frames 1, 12, 24; then render 24 frames at 720p."

A 4-frame preview rendered by JANCTION Render (a rotating red cube)
The 4-frame preview Claude.ai got from that prompt: 4 GPU seconds. The 24-frame 720p MP4 that followed took 58 GPU seconds (2026-10-01, RTX PRO 6000).

How long it takes

jobGPU timenote
preview, 4 frames (720p split in 4 tiles, 16 samples, denoised)about 4 sround trip about 8 s; the image comes back inline
final, 24 frames at 1280x720, 64 samples → MP4about 60 smeasured 2026-09-30 / 10-01 (59 s and 58 s)
final, one 1920x1080 frame at 128 samplesabout 11 sthe estimator's baseline; heavy scenes take longer

These are individual observations of simple scenes (2026-09-30 to 10-01 UTC, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96 GB). The 4-GPU-second example used a rotating cube, four 640x360 tiles and 16 samples. GPU time includes Blender startup and scene loading; it is distinct from completion time including queue and transfers. See conditions, round trips and reproduction files. These are not speed guarantees.

render_estimate gives the number before rendering; once a scene has been rendered once, the estimate uses that scene's own measurement.

Connect

fromhow
A file on this computer (.blend, 3D file)Drop it at /upload to get a 12-hour link (no sign-up), then tell the AI "Render this file: <link>". From inside Blender use the add-on; from a terminal the CLI uploads directly
Claude.ai (web / desktop / mobile)Plugins (older menu: Customize → Connectors) → Add custom connector → https://render.janction.jp/mcp. "Connect" opens a page that creates a free key
Claude Team / EnterpriseAn owner adds https://render.janction.jp/mcp under Organization settings → Connectors; each member then presses Connect with their own account. In organizations that turn Skills off, Claude Desktop also blocks plugins from November 30 (use the connector URL, or ask the admin to turn Skills on)
ChatGPTPlugins (left sidebar) → Add → Create plugin → https://render.janction.jp/mcp (OAuth). Attach your .blend to the chat. On Business workspaces an admin adds it and shares it
Claude Codeclaude mcp add --transport http janction-render https://render.janction.jp/mcp, then /mcp to authenticate
Claude Code on the web (claude.ai/code)The same connector as Claude.ai: Plugins (older menu: Customize → Connectors) → Add custom connector → https://render.janction.jp/mcp. Works without changing the session's network settings (guide)
Codexcodex mcp add janction-render --url https://render.janction.jp/mcp (the sign-in page opens right away; if Codex later says the server needs login, run codex mcp login janction-render; with a key, add --bearer-token-env-var JANCTION_RENDER_API_KEY)
Cursor / VS CodeAdd to Cursor Add to VS Code Add to LM Studio
The button opens the editor; confirm and it is added. Manual: Streamable HTTP at https://render.janction.jp/mcp
Antigravity CLI / Gemini CLIagy plugin install https://github.com/JasmyLab-JANCTION/janction-render or gemini extensions install https://github.com/JasmyLab-JANCTION/janction-render (the MCP server and the workflow skill)
Kimi Code (Kimi K3 and other models)/plugins install https://github.com/JasmyLab-JANCTION/janction-render (the MCP server and the workflow skill), or add https://render.janction.jp/mcp as a url entry under mcpServers in ~/.kimi-code/mcp.json and run /mcp-config login janction-render (OAuth). Guide
OpenCodeAdd https://render.janction.jp/mcp under mcp in ~/.config/opencode/opencode.json with "type": "remote", then opencode mcp auth janction-render (OAuth). Guide
Qwen Codeqwen extensions install https://github.com/JasmyLab-JANCTION/janction-render (the MCP server and the workflow skill), or qwen mcp add --transport http janction-render with https://render.janction.jp/mcp (stored as httpUrl). Guide
GitHub Copilot CLIcopilot mcp add --transport http janction-render with https://render.janction.jp/mcp, then /mcp auth janction-render in copilot (OAuth). With the skill: copilot plugin marketplace add JasmyLab-JANCTION/janction-render, then copilot plugin install janction-render@janction-render. Guide
DeepSeek Harness (dsh)One @deepseek-ai/dsh-mcp-client entry: transport: streamable-http, url: https://render.janction.jp/mcp, headers Authorization: Bearer with a key from POST /v1/keys (no OAuth). DeepSeek guide (API function calling too)
Grok / Perplexity / Le ChatAdd a custom (remote MCP) connector with https://render.janction.jp/mcp in each app's Connectors settings (Grok: Plugins → New connector → Custom, guide; Perplexity: Settings → Connectors → Custom → Remote, Pro / Max / Enterprise, OAuth; Le Chat: workspace admin, OAuth or a Bearer key)
Gemini app (gemini.google.com)On the web: Settings → Connected Apps → Custom apps, add the MCP URL https://render.janction.jp/mcp (web only, English interface, with activity saving on). Sign-in is the OAuth consent page
ManusConnectors → Create → Add MCP by URL with https://render.janction.jp/mcp (remote MCP only; OAuth or a Bearer key)
MiniMax (MiniMax Code)One streamable-http entry in the MCP settings: url https://render.janction.jp/mcp, and Authorization: Bearer jr_… in headers
Notion (Notion Agent)Business / Enterprise: Settings → Connections → Discover → Add Custom MCP → https://render.janction.jp/mcp, then sign in (the button shows when a workspace admin allows custom MCP servers)
JetBrains (AI Assistant, Junie)AI Assistant: Settings → Tools → AI Assistant → Model Context Protocol (MCP) → Add → connect over HTTP with {"mcpServers": {"janction-render": {"url": "https://render.janction.jp/mcp"}}}. Junie: the same url in ~/.junie/mcp/mcp.json (Authorize opens the browser), or "headers": {"Authorization": "Bearer jr_…"}
Kiro (IDE, CLI) / Amazon Q Developer CLIKiro: "url": "https://render.janction.jp/mcp" under mcpServers in ~/.kiro/settings/mcp.json (in the IDE, open a folder: MCP SERVERS appears in the Kiro panel and the authorization page opens; in the CLI, /mcp auth janction-render). With a key: "headers": {"Authorization": "Bearer ${JANCTION_RENDER_API_KEY}"} (add the variable to Mcp Approved Env Vars). Amazon Q Developer CLI: "type": "http" and the same url in the agent configuration, then /mcp to sign in
Microsoft Copilot StudioTools → Add a tool → New tool → Model Context Protocol → Server URL https://render.janction.jp/mcp. Authentication: OAuth 2.0 with Dynamic discovery, or API key (Header, name X-API-Key, value jr_…). Set the agent's Settings → Generative AI → Orchestration to Yes (generative orchestration), or it will not use the tool
LM Studio (0.3.17 or later)Program tab → Install → Edit mcp.json: {"mcpServers": {"janction-render": {"url": "https://render.janction.jp/mcp", "headers": {"Authorization": "Bearer jr_…"}}}}
Cherry StudioSettings → MCP Servers → Add → Import from JSON: {"mcpServers": {"janction-render": {"type": "streamableHttp", "baseUrl": "https://render.janction.jp/mcp"}}}. Turn it on and press "Sign-in required." to open the consent page in the default browser. If no tools show after signing in, turn the server off and on again (with a key: "headers": {"Authorization": "Bearer jr_…"})
Windsurf, Cline, Goose, Amp, Auggie, Warp, Zed, Roo Code, Continue and other MCP clientsStreamable HTTP at https://render.janction.jp/mcp; OAuth or Authorization: Bearer <api key>
Claude Desktop and other stdio-only clients (npm)npx -y janction-render (in the client's mcpServers: {"command": "npx", "args": ["-y", "janction-render"]}). The npm package janction-render bridges stdio to https://render.janction.jp/mcp with OAuth; --key jr_… for a key you have, --local for the Python stdio server through uv
Terminal (stdio, sends local files)claude mcp add janction-render -e JANCTION_RENDER_SERVER=https://render.janction.jp -- uvx --from janction-render janction-render-mcp
Inside Blender (for people)The add-on; extension repository https://render.janction.jp/extensions/index.json. Maya / Houdini / Cinema 4D: submit tools

Free limits

Pricing (since 2026-10-08): Each new key gets 100 JPY of GPU time free with no card (1,000 GPU-seconds, 14 days) for previews and finals, and 400 JPY more once a card is saved on a Stripe page that charges nothing (once per card): up to 500 JPY in all. After it is used up or expires, previews are free up to 2 GPU-minutes per key per day and final renders are paid from prepaid credit at 0.1 JPY per GPU-second. Top-ups (paid) start at 100 JPY. The first top-up during the 14-day welcome period counts 1.5x, with the bonus capped at 500 JPY (100 JPY buys 150 JPY of credit). Top-up bonuses are separate from the welcome credit and the card-saving credit. Payment is on a Stripe Checkout page. One full welcome credit per network every 30 days; further keys from the same network use what is left of that network's credit. A job without enough credit gets HTTP 402 payment_required with a checkout_url. Final renders go up to 3,000 frames per instruction (automatically split internally under one job id), 1080p (1920x1080) on free time and up to 4K (3840x2160) on prepaid credit. The free preview time resets at 00:00 UTC (09:00 JST). No sign-up: connecting creates a key. Inputs and results are deleted 24 hours after last use (normally 7 days for keys that have paid 500 JPY or more in total) and never used for training.

Without a connector

If you cannot add a connector to your AI (plan or administrator settings), the same result is one step away: send the prompt below to the AI and paste the code it returns into the paste-and-render page. The preview needs no payment.

Write a Blender 5.0 bpy script that builds <describe your scene here>. Add a camera and set scene.camera, and add one light. Do not call render and do not change render settings. Reply with the code only.

To fix something, give the AI the preview's notes (too dark, cut off, no light) and paste the rewritten script again.

Checked per app

A model name alone does not decide whether it connects: the app, its plan, administrator settings and tool permissions do. "verified" means external users rendered from that app.

appmodelconnectauthneedschecked scenariostatuschecked
Claude.ai (web / desktop / mobile)Claude (the app's current models)Plugins (older menu: Customize > Connectors) > custom connectorOAutha plan with custom connectorsagent-written bpy and .blend by link: preview, fix, finalverified2026-10-11
ChatGPTthe app's current modelsPlugins > Add > Create pluginOAutha plan that allows custom plugins (Business: an admin adds it)sameverified2026-10-28
Claude CodeClaudeclaude mcp add --transport http (or the plugin)OAuth-bpy from the terminal, local files through the stdio packageverified2026-10-11
Codex (CLI / IDE)the app's current modelscodex mcp add (or the plugin)OAuth-sameverified2026-10-11
Grok (grok.com)GrokConnectors > CustomOAuth-preview and finalverified2026-10-11
Notion AgentNotion's modelsSettings > Connections > Add Custom MCPOAuthBusiness / Enterprise; a workspace admin allows custom MCPpreview and finalverified2026-10-11
OpenCodeanyStreamable HTTP MCPOAuth or Bearer key-preview and finalverified2026-10-11
Cursorthe editor's modelsmcp.json (one-click button on the guide)OAuth-preview, a 10-frame test, then a 792-frame finalverified2026-10-27
VS Code / Windsurfthe editor's modelsmcp.json (one-click button for VS Code)OAuth--not verified by external use yet2026-10-11
Gemini CLI / Antigravity CLIGeminiextension / plugin installOAuthGoogle sign-in-installed by the operator; not verified by external use yet2026-10-11
Kimi Code (CLI / VS Code)Kimi K3 and others/plugins install (or mcp.json)OAutha Kimi Code login-not verified by external use yet2026-10-27
Qwen CodeQwen and othersqwen extensions install (or qwen mcp add)OAuth--not verified by external use yet2026-10-27
GitHub Copilot CLIthe plan's modelscopilot mcp add (or the plugin)OAuth or a key headera Copilot plan; the organization allows the CLIpreview of the sample (CLI 1.0.107)verified2026-11-09
JetBrains AI Assistant / Juniethe IDE's modelsMCP settings (AI Assistant) or mcp.json (Junie)OAuth (Junie) or a key header (Junie)--not verified by external use yet2026-11-09
Kiro (IDE / CLI) / Amazon Q Developer CLIthe app's models~/.kiro/settings/mcp.json (url)OAuth or a key headera folder open in the IDE (the MCP panel shows with a workspace)preview of the sample (Kiro IDE 1.4.121, OAuth)checked by the operator; not verified by external use yet2026-11-09
Microsoft Copilot Studiothe agent's modelsTools > Add a tool > Model Context ProtocolOAuth 2.0 (dynamic discovery) or API key header X-API-Keygenerative orchestration on the agentpreview of the sample (OAuth dynamic discovery, generative orchestration on)checked by the operator; not verified by external use yet2026-11-09
LM Studiolocal modelsmcp.json (url and headers)a key headerLM Studio 0.3.17 or later-not verified by external use yet2026-11-09
Cherry Studiothe configured modelsSettings > MCP Servers > Add > Import from JSON (streamableHttp)OAuth or a key headerturn the server off and on after signing inconnects over OAuth and lists the tools (2.1.10)not verified by external use yet2026-11-09
Gemini app (gemini.google.com)GeminiSettings > Connected Apps > Custom apps (MCP URL, web)OAuththe web app in English with activity saving on-not verified by external use yet2026-11-09
ManusManusConnectors > Create > Add MCP by URLOAuth or a Bearer key--not verified by external use yet2026-11-09
MiniMax CodeMiniMaxMCP settings (streamable-http url and headers)a Bearer key--not verified by external use yet2026-11-09
Any AI without a connectoranynone: the AI writes the bpy script, the person pastes it at /startnone (a key is created in the browser)-preview, fix, final in the browseravailable2026-10-11

FAQ

Can an AI generate a 3D scene from text with this?

The agent generates the scene as a bpy script (Blender's Python), and JANCTION Render renders it. There is no text-to-3D model inside; the quality comes from the agent's script and the preview loop.

How many previews does a typical scene take?

Two or three. Each is a few GPU seconds (4 s for a simple scene), so even ten previews stay far inside the welcome credit, and after it inside the free previews (up to 2 GPU-minutes a day).

Can the agent see the preview?

Yes. In MCP the preview comes back as an image the agent can look at (and as an interactive panel in clients with MCP Apps); over HTTP it is a PNG link.

Related

Overview · All guides · llms.txt