JANCTION Render

What is JANCTION Render?

JANCTION Render is a cloud GPU render farm for Blender that AI agents call as an MCP server or HTTP API. Send a .blend file, a bpy script or a 3D file (glTF / FBX / USD), preview in seconds, and get frames or an MP4 back from JANCTION's NVIDIA GPUs. Every preview also returns a verdict on what is wrong, such as too dark, blown out, no light or an object out of frame, and the bpy code to fix it, so the agent can correct the scene before the final render. No local Blender or GPU is needed. A 100 JPY welcome credit for a new key (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), then 0.1 JPY per GPU-second (previews free up to 2 GPU-minutes a day); operated by JasmyLab Inc.

AI agents can render Blender projects on JANCTION GPUs without running Blender locally. Works from Claude, ChatGPT, Claude Code, Codex, Cursor or any MCP client: an MCP server (remote and stdio), an HTTP API and a CLI.

JANCTION Render is a separate product from SmartRender (JasmyLab's distributed rendering for people at a desktop) and is not affiliated with Render.com or the Render Network. The numbers, as of 2026-10-08 (the primary source is https://render.janction.jp/facts.json): a 100 JPY welcome credit (1,000 GPU-seconds, 14 days, no card) for a new key, one per network every 30 days (further keys on the same network share what is left), plus 400 JPY more once a card is saved on a page that charges nothing (once per card); after that, previews are free up to 2 GPU-minutes a day and finals cost 0.1 JPY per GPU-second from prepaid credit (paid use started on 2026-10-08); finals up to 3,000 frames per instruction (run as parts of 240 under one job id); 1080p on free time, up to 4K on prepaid credit; inputs and results deleted 24 hours after last use (7 days, normally, for keys that have paid 500 JPY or more in total).

Official site: https://render.janction.jp · MCP endpoint: https://render.janction.jp/mcp · Fact sheet: https://render.janction.jp/facts

Try it from a terminal in a few seconds, free and with nothing to sign up for (needs uv; or pip install janction-render and drop the uvx):

uvx janction-render try                      # renders a sample scene and shows the free GPU time of your new key
uvx janction-render preview scene.blend      # your .blend, bpy script or 3D file
uvx janction-render render scene.blend --frames 1-48 --output mp4 --wait
uvx janction-render turntable model.glb      # a 360-degree turntable MP4 in one command (a path or an https link)
uvx janction-render shots model.glb --transparent   # four product shots (PNG)

What one command makes (real outputs)

Rendered on 2026-10-10 on the production GPU (NVIDIA RTX PRO 6000 Blackwell) from Khronos's WaterBottle sample (public domain, CC0); copy any line and it runs as is. Prices are at 0.1 JPY per GPU-second and are covered by the welcome credit of a new key.

	Command	GPU time	Price
	uvx janction-render shots https://github.com/KhronosGroup/glTF-Sample-Assets/blob/main/Models/WaterBottle/glTF-Binary/WaterBottle.glb	13.9 s	about 1.4 JPY
	uvx janction-render turntable https://github.com/KhronosGroup/glTF-Sample-Assets/blob/main/Models/WaterBottle/glTF-Binary/WaterBottle.glb	85.2 s	about 8.5 JPY
	uvx janction-render render https://render.janction.jp/samples/ex12_logo_ring_spin.py --frames 1-48 --size 1280x720 --engine eevee --output mp4 --wait	54.1 s	about 5.4 JPY

The shots are four 1920x1080 PNGs; the turntable and the animation are MP4s (shown here as GIFs). Your own .glb, .blend or bpy script goes in place of the link. More, with their scripts: https://render.janction.jp/examples

Keep 3D files in a repository? Add a button that lets anyone render them from their own AI or terminal (links to file pages on GitHub and Hugging Face work as they are; make one at https://render.janction.jp/go):

[![Render on JANCTION](https://render.janction.jp/badge.svg)](https://render.janction.jp/go?scene=https://github.com/you/repo/blob/main/model.glb)

Guides on the official site

• How to render Blender without a GPU
• From an agent: Claude Code + Blender · Codex + Blender · Cursor + Blender · ChatGPT + Blender · Blender MCP server for rendering
• Best render farms for AI agents (2026): JANCTION, BlendSwap, Sceneplane, your own worker on Modal or RunPod and a local Blender MCP side by side
• Headless Blender as an API · glTF, GLB or FBX to MP4 · 360-degree product video · What a Blender render costs
• Examples with their bpy scripts · Compare · 日本語のご案内

Why use it?

• No GPU, no Blender install. The scene can be a bpy script the agent writes, a .blend, or a 3D file; Blender runs on our GPUs.
• Built for agents. Preview first (1-4 frames in a few GPU seconds), an estimate with time and cost before the final, structured errors with the fix, spending caps per key, safe retries with Idempotency-Key.
• 3D files and turntables. glTF / GLB, FBX, USD, OBJ, STL, PLY, Alembic are imported into an empty scene with a camera and HDRI lighting; orbit=True makes a turntable. Fixed-price outcomes: POST /v1/outcomes/turntable and /product-shot.
• Honest limits. a 100 JPY welcome credit for a new key with no card (14 days; one per network every 30 days) and 400 JPY more after saving a card that is not charged, then 0.1 JPY per GPU-second (previews free up to 2 GPU-minutes a day), finals up to 3,000 frames per instruction as parts of 240 (1080p on free time, up to 4K on prepaid credit), on JANCTION's own GPUs (NVIDIA RTX PRO 6000 Blackwell); when they are busy jobs wait in a queue and the ETA says so. Inputs and results are deleted 24 hours after last use (7 days, normally, for keys that have paid 500 JPY or more in total) and never used for training.
• Not Render.com. Same word, different product.

Quick start

Claude.ai / ChatGPT (nothing to install). Settings → Connectors → add https://render.janction.jp/mcp → Connect (a free key is created on the consent page). Then paste:

Render a preview of https://render.janction.jp/samples/cube_scene.py

A 4-frame preview comes back in a few seconds. A file on your computer: drop it at https://render.janction.jp/upload to get a 12-hour link, then "Render this file: ".

Claude Code.

claude mcp add --transport http janction-render https://render.janction.jp/mcp

Then /mcp to authenticate and ask: "Build a small street scene in Blender with a camera fly-through and render a preview with janction-render."

HTTP (any language).

curl -s -X POST https://render.janction.jp/v1/keys                                   # -> {"api_key": "jr_..."}  no sign-up
curl -s -X POST https://render.janction.jp/v1/files -H "X-API-Key: $KEY" -F "file=@scene.blend"   # -> {"scene_id": "f_..."}
curl -s -X POST https://render.janction.jp/v1/jobs -H "X-API-Key: $KEY" -H "Content-Type: application/json" \
  -d '{"scene_id":"f_...","kind":"preview","frames":[1]}'                             # -> job_id; then GET /v1/jobs/{job_id}

From your own agent or app

Python (pip install janction-render; the first call creates a free key and keeps it in ~/.janction-render.json):

from janction_render.client import Client

c = Client()                                   # or Client(api_key="jr_...")
scene = c.upload("scene.py")                   # a bpy script, a .blend, or a .glb / .fbx / .usd / .obj file
job = c.submit(scene["scene_id"], kind="final", frame_start=1, frame_end=48, output="mp4")
job = c.wait(job["job_id"], timeout=1800)
print(job["status"], c.download(job["job_id"], "out", only="mp4"))

Pay per use with prepaid credit; no subscription is needed (prices: https://render.janction.jp/pricing). Instead of polling, pass notify_url (https) to POST /v1/jobs and get one JSON POST when the job finishes (job.done / job.failed, with download links); an Idempotency-Key header makes retries safe. OpenAPI: https://render.janction.jp/openapi.json A complete app example (a user uploads a GLB, the app orders fixed-price product shots, a webhook stores the images): https://render.janction.jp/blender-render-api#embed

Agent frameworks connect to the remote MCP server at https://render.janction.jp/mcp with the key as a bearer token (curl -s -X POST https://render.janction.jp/v1/keys gives one, no sign-up).

OpenAI Agents SDK (keep the longer timeout: render_preview waits for the image, longer than the 5-second default):

import asyncio
from agents import Agent, Runner
from agents.mcp import MCPServerStreamableHttp

async def main():
    async with MCPServerStreamableHttp(name="janction-render", client_session_timeout_seconds=120,
                                       params={"url": "https://render.janction.jp/mcp",
                                               "headers": {"Authorization": "Bearer jr_..."}}) as render:
        agent = Agent(name="3D artist", instructions="Render Blender scenes with janction-render.", mcp_servers=[render])
        result = await Runner.run(agent, "Render a preview of https://render.janction.jp/samples/cube_scene.py")
        print(result.final_output)

asyncio.run(main())

LangChain (pip install langchain-mcp-adapters):

import asyncio
from langchain_mcp_adapters.client import MultiServerMCPClient

async def main():
    client = MultiServerMCPClient({"janction-render": {"url": "https://render.janction.jp/mcp", "transport": "streamable_http",
                                                       "headers": {"Authorization": "Bearer jr_..."}}})
    tools = await client.get_tools()           # render_preview, render_final, render_status, render_download, ...
    print([t.name for t in tools])

asyncio.run(main())

OpenAI Responses API (remote MCP; render_final keeps asking for approval because it spends credit):

from openai import OpenAI

client = OpenAI()
resp = client.responses.create(
    model="gpt-5",
    tools=[{"type": "mcp", "server_label": "janction_render", "server_url": "https://render.janction.jp/mcp",
            "authorization": "jr_...",
            "require_approval": {"never": {"tool_names": ["scene_info", "render_preview", "render_estimate",
                                                         "render_status", "render_download"]}}}],
    input="Render a preview of https://render.janction.jp/samples/cube_scene.py and give me the image link.",
)
print(resp.output_text)

Anthropic Messages API (MCP connector; fallbacks="default" reruns a declined request on a fallback model):

import anthropic

client = anthropic.Anthropic()
response = client.beta.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    betas=["mcp-client-2025-11-20", "server-side-fallback-2026-07-01"],
    fallbacks="default",
    mcp_servers=[{"type": "url", "url": "https://render.janction.jp/mcp", "name": "janction-render",
                  "authorization_token": "jr_..."}],
    tools=[{"type": "mcp_toolset", "mcp_server_name": "janction-render"}],
    messages=[{"role": "user", "content": "Render a preview of https://render.janction.jp/samples/cube_scene.py"}],
)
for block in response.content:
    if block.type == "text":
        print(block.text)

Google ADK:

import os
from google.adk.agents import Agent
from google.adk.tools.mcp_tool import MCPToolset, StreamableHTTPConnectionParams

agent = Agent(
    name="renderer",
    model="gemini-2.5-flash",
    instruction="Render Blender scenes with JANCTION Render: preview first, ask before a final render.",
    tools=[MCPToolset(connection_params=StreamableHTTPConnectionParams(
        url="https://render.janction.jp/mcp",
        headers={"Authorization": "Bearer " + os.environ["JANCTION_RENDER_API_KEY"]}))],
)

Pydantic AI:

from pydantic_ai import Agent
from pydantic_ai.mcp import MCPToolset

render = MCPToolset("https://render.janction.jp/mcp", headers={"Authorization": "Bearer jr_..."})
agent = Agent("anthropic:claude-opus-5-5", toolsets=[render])

Connect the AI apps on this computer in one command, and tell every coding agent in a project when and how to render:

uvx janction-render connect             # lists Claude Code, Codex, Gemini CLI, Cursor, Windsurf, VS Code and the step for each
uvx janction-render connect claude-code # applies one (or: connect all --yes)
uvx janction-render init                # writes a short section to AGENTS.md and adds the server to .mcp.json

init touches CLAUDE.md, GEMINI.md, .cursor/ and .vscode/ only when they already exist (or with --claude, --cursor, --vscode); it keeps everything else in those files and backs up JSON it changes.

The local MCP server (stdio) also offers prompts that show up as slash commands, for example /mcp__janction-render__turntable_video in Claude Code: try_janction_render, render_my_scene, turntable_video, product_shots, scene_from_description, blend_to_mp4.

MCP

Remote (Streamable HTTP, OAuth 2.1 with dynamic client registration; a raw key also works as Authorization: Bearer jr_...)

Remote tools take scene_script (bpy code as text), scene_url (an https link to a .blend, .py or a 3D file, or a link from https://render.janction.jp/upload) or scene_id, plus asset_urls for textures or glTF .bin files. Results come back as an inline image plus download links that need no key and work for about 24 hours (render_download(only="mp4") for just the video).

Claude Code plugin (the remote connector plus a skill with the workflow):

/plugin marketplace add JasmyLab-JANCTION/janction-render
/plugin install janction-render@janction-render

Stdio (sends local files)

The package is on PyPI as janction-render (and the npm launcher janction-render runs it with npx -y janction-render --local when uv is installed).

# Claude Code (uvx runs it without a global install)
claude mcp add janction-render -e JANCTION_RENDER_SERVER=https://render.janction.jp -- uvx --from janction-render janction-render-mcp

# or with pip / pipx
pip install janction-render
claude mcp add janction-render -e JANCTION_RENDER_SERVER=https://render.janction.jp -- janction-render-mcp

Codex: add to ~/.codex/config.toml

[mcp_servers.janction-render]
command = "uvx"
args = ["--from", "janction-render", "janction-render-mcp"]
env = { JANCTION_RENDER_SERVER = "https://render.janction.jp" }

A temporary API key is issued automatically on first use and cached in ~/.janction-render.json; set JANCTION_RENDER_API_KEY to pin one (quota and, later, credit belong to the key). The same key can be used from the remote connector: paste it on the connect page.

Tools

Options on render_preview / render_final: environment (studio, sunset, overcast, night, compare on previews; environment_strength, environment_visible=False for a flat grey backdrop with HDRI lighting), orbit / orbit_frames / orbit_elevation (turntable), blender ("5.2"), and assets (stdio: local files sent with the scene, found in the script through os.environ["JR_ASSETS_DIR"]; a .blend's external files are collected automatically with pip install janction-render[blend]) or asset_urls (remote).

CLI: janction-render inspect|preview|render|status|download|share|unshare|cancel|jobs|balance|topup|limits|outcomes|info.

API

POST /v1/keys                                   -> {api_key}         (header X-API-Key afterwards)
POST /v1/files  multipart "file" (.blend|.py|.glb|.fbx|.usd|.obj|...)   -> {scene_id}
POST /v1/files/{id}/assets  multipart "files"   textures, glTF .bin ...   GET /v1/files/lookup?sha256=  reuse an upload
POST /v1/estimate {kind, frames|frame_start/frame_end, width, height, samples, scene_id?} -> seconds, wall_seconds, human, quota, currency, expires_at
POST /v1/jobs   {scene_id, kind: info|preview|final, frames|frame_start/frame_end, width, height, samples, camera, fps, output (png|exr|mp4|webm|prores|gif|webp), transparent, notify_url (https), engine (cycles|eevee),
                 environment, environment_strength, environment_visible, blender, orbit, orbit_frames, orbit_elevation}   header Idempotency-Key makes retries safe
GET  /v1/jobs/{id}      status, progress, eta, artifacts[], cost, warnings, info, failure (code + fix)    DELETE /v1/jobs/{id}  cancel
GET  /v1/jobs/{id}/artifacts/{name}             PNG / MP4
GET  /v1/jobs/{id}/archive?only=png             every file of the job in one zip (frame sequences: one request instead of hundreds)
GET  /v1/outcomes       fixed-price outcomes;  POST /v1/outcomes/turntable {scene_id, size, frames}   POST /v1/outcomes/product-shot {scene_id, size, transparent?}
POST /v1/try    JSON {scene_script | scene_url, frames?, width?, height?, environment?} (no key) -> first 720p preview + an API key to continue, in one response (1 per network per 24h)
POST /v1/drops  multipart "file" (no key)       -> a 12-hour https link to pass as scene_url (the /upload page uses this)
POST /v1/jobs/{id}/share {title?, note?, include_script?, listed?} -> {share_id, url}   DELETE /v1/jobs/{id}/share   GET /v1/shares
GET  /r/{share_id}  public page (no key)   GET /gallery
GET  /v1/assets/search?q=&kind=models|textures|hdris   CC0 assets (Poly Haven) -> spec polyhaven:<id>
POST /v1/files/{id}/assets/urls {urls: ["https://...", "polyhaven:<id>"]}   fetch assets server-side
GET  /v1/me  (quota, balance, limits)   POST /v1/me/limits {job_yen?, day_yen?}   spending caps per key
POST /mcp                                       remote MCP (Streamable HTTP; Bearer api key or OAuth)
GET  /.well-known/oauth-protected-resource/mcp  OAuth discovery
POST /v1/billing/checkout {amount_yen} -> {checkout_url}   POST /v1/billing/sync   GET /v1/ledger
GET  /openapi.json  /llms.txt  /llms-full.txt  /ja/llms.txt  /facts.json  /version.json  /capabilities.json  /pricing.json  /status.json

A job that goes past the free GPU time without enough credit gets 402 payment_required with checkout_url (show it to the user), topup_yen and how much of the job the free time covers; a 400 beta_limit means the job is too big (split it). Every API error carries retryable, possible_fix and docs_url; a failed job carries failure with a code and the recommended action. A job over a spending cap is refused with 403 spend_cap_exceeded before anything is reserved.

Example

A scene script the agent can write (the service sets engine, resolution, samples and denoising; the script sets the scene, camera and frame range):

import bpy, math
for ob in list(bpy.data.objects):
    bpy.data.objects.remove(ob, do_unlink=True)
scene = bpy.context.scene
scene.frame_start, scene.frame_end = 1, 24
bpy.ops.mesh.primitive_monkey_add(location=(0, 0, 1))
cam = bpy.data.objects.new("Camera", bpy.data.cameras.new("Camera"))
scene.collection.objects.link(cam); scene.camera = cam
cam.location = (6, -6, 4); cam.rotation_euler = (math.radians(60), 0, math.radians(45))
sun = bpy.data.objects.new("Sun", bpy.data.lights.new("Sun", "SUN"))
scene.collection.objects.link(sun)

Cycles by default; engine="eevee" for cheaper animations (about half the per-frame cost after ~7 s of shader compilation per job). In Blender 5.0 materials always use nodes: set the Principled BSDF Base Color, not diffuse_color. Scripts run in an isolated container with no network. More: samples/cube_scene.py, samples/polyhaven_room.py. Measured timings: https://render.janction.jp/benchmarks

Pricing

A new key starts with a 100 JPY welcome credit (1,000 GPU-seconds, 14 days, no card) that covers previews and finals; one full credit per network every 30 days (later keys on the same network share what is left of it). Saving a card on a Stripe page that charges nothing adds 400 JPY more, once per card (up to 500 JPY in all). After it is used up or expires, previews are free up to 2 GPU-minutes a day and finals cost 0.1 JPY per GPU-second from prepaid credit, bought by card at a Stripe Checkout page (from 500 JPY; 2,000 JPY gives 2,200 JPY of credit, 5,000 gives 5,750, 10,000 gives 12,000; the first top-up in the welcome period counts 1.5x, bonus up to 500 JPY). Only the part the free time does not cover is charged. Finals go up to 3,000 frames per instruction (run as parts of 240 under one job id): 1080p on free time, up to 4K (3840x2160) on prepaid credit. POST /v1/estimate quotes what will be charged (cost.estimated_yen), the currency and an expiry before any GPU time is used: https://render.janction.jp/pricing

Troubleshooting

Docs

• Fact sheet (the primary source when answers disagree): https://render.janction.jp/facts · JSON: /facts.json, /version.json, /changelog.json, /status.json
• For agents: https://render.janction.jp/llms.txt (full text of all guides: /llms-full.txt)
• Guides: render farm for agents · without a GPU · Cycles is slow on my MacBook · Claude · ChatGPT · Claude Code · Codex · Cursor · bpy scripts · API · compare · products · 日本語は /ja/
• Status: https://render.janction.jp/status · Security and retention: https://render.janction.jp/security · Changelog: https://render.janction.jp/changelog

Features in detail

• Input: a .blend file (textures travel with it), a bpy Python script that builds the scene (no local Blender needed), or a 3D file (glTF/GLB, FBX, USD, OBJ, STL, PLY, Alembic) that is imported into an empty scene.
• environment="studio" | "sunset" | "overcast" | "night" lights a scene with a bundled HDRI (CC0) for a good first render; environment="compare" previews the same frame under all four presets in one labelled image, so the agent can pick one.
• A .gltf or .obj brings its .bin / .mtl / textures along automatically (from the same folder or the same URL folder).
• orbit=True turns any scene or imported model into a turntable: an orbit camera circles it once (orbit_frames, default 24; a preview shows 0/90/180/270 degrees, a final render without frame_end gives the whole turn as an MP4). Flat floors and walls are ignored when framing; orbit_target (object name or x,y,z) and orbit_distance (multiplier) adjust the shot.
• While the GPU is lent to another workload, tools answer right away with the queued job and a wait estimate (how long it has been out, how long it usually stays out, how much longer to expect; eta.gate) instead of waiting.
• While a job renders, render_status counts finished frames inside the running chunk, so the ETA updates every few seconds.
• In hosts that support MCP Apps (Claude web and desktop, among others), previews, progress and download links also appear as an interactive panel in the chat: the image, a progress bar, "Open MP4", and preset buttons after an environment="compare" preview. Both the remote connector and the stdio package ship the panel (janction_render/mcp_app.html); set JR_MCP_APPS=0 to turn it off.
• CC0 assets by name: asset_search("wooden table") finds Poly Haven models, textures and HDRIs; pass polyhaven:<id> in asset_urls (remote) or assets (stdio) and the files land in JR_ASSETS_DIR/<id>/ on the GPU, ready to import (bpy.ops.import_scene.gltf(filepath=os.path.join(os.environ['JR_ASSETS_DIR'], 'wooden_table_02', 'wooden_table_02_1k.gltf'))). Scene scripts can import jr_assets for one-liners: jr_assets.import_model('wooden_table_02'), jr_assets.apply(floor, jr_assets.material('wood_floor_deck', scale=2)) (diff / normal / roughness / metal wired into a Principled BSDF), jr_assets.world_hdri('studio_small_09'). See samples/polyhaven_room.py.
• A render you choose to share (render_share) gets a public page (/r/<id>, OGP for X and Discord) that stays until you unshare it.
• Blender 5.0, Cycles on GPU (blender="5.2" selects Blender 5.2 when the worker has it; render_info lists what is available).

Blender add-on and DCC tools (for people, not agents)

• Blender extension (addons/blender/, Blender 4.2+): a JANCTION Render panel under Properties → Render with Preview, Estimate and Final. A copy of the open file and the textures it references with relative paths are uploaded; results land next to the .blend. Install from the extension repository https://render.janction.jp/extensions/index.json (Preferences → Get Extensions → Repositories → + → Add Remote Repository), or install the zip from https://render.janction.jp/extensions/janction_render-0.1.2.zip (python scripts/build_blender_addon.py builds it from this repository). Guide: https://render.janction.jp/blender-addon
• Maya, Houdini, Cinema 4D (integrations/): export to USD / FBX / Alembic, upload with textures, render with Cycles, open the result; Preview, Final (MP4) and Turntable. Standard-library Python only; jr_submit.py is the shared client. Guide: https://render.janction.jp/maya-houdini-cinema4d

Self-hosting

The server (FastAPI + SQLite, with the remote MCP endpoint) and the worker (Blender in disposable Docker containers, --network none --cap-drop ALL) live in the internal repository and are not part of this package yet. This repository holds the client side: stdio MCP server, CLI, HTTP client, samples, the Claude Code plugin, the Blender add-on and the DCC tools.

MCP registry

This server is listed in the official MCP registry as io.github.JasmyLab-JANCTION/janction-render (remote: https://render.janction.jp/mcp).

mcp-name: io.github.JasmyLab-JANCTION/janction-render

Security

Vulnerability reports: see SECURITY.md. Retention, isolation and the external tests are summarised on the security page.

License

MIT (see LICENSE) for the package, the MCP servers, the CLI and the DCC tools. The Blender extension under addons/blender/ is GPL-3.0-or-later, as Blender requires for add-ons. Operated by JasmyLab Inc.