JANCTION Render

Render Blender scenes on a cloud GPU from AI agents (Claude, ChatGPT, Codex, Cursor). Send a .blend, a bpy script or a glTF/FBX/USD file; get a 4-frame preview, a time estimate, then final frames or an MP4. Remote MCP with OAuth or stdio via pip. Free beta: 10 GPU-minutes per key per day.

Hosted MCP Server

npx add-mcp 'https://render.janction.jp/mcp'

Installs into Claude Code, Codex, Cursor and more

Documentation

JANCTION Render

JasmyLab-JANCTION/janction-render MCP server

A cloud GPU render farm for Blender, built for AI agents. Render Blender scenes on JANCTION GPUs from Claude, ChatGPT, Claude Code, Codex, Cursor or any MCP client: an MCP server (remote and stdio) plus a CLI. Use it to render Blender without a GPU, or when rendering locally is slow.

An execution layer for agent-authored scenes and existing .blend files: edit in your agent or local Blender, then send the scene to cloud GPUs for rendering.

  • 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.
  • scene_info reads the scene without rendering (cameras, frame range, missing files).
  • render_preview returns 1-4 low-cost frames tiled in one image within seconds, so the agent can look, fix, and repeat.
  • render_estimate says how long a render will take ("about 3 minutes") and whether it fits today's free quota.
  • render_final renders the frames on GPUs and joins them into an MP4; render_status reports the remaining time.
  • Free beta: no charges. Each key gets 10 GPU-minutes per day; a final render is up to 240 frames at 1080p. Paid plans (per GPU second, prepaid credit) will be announced on the service page before they start.
  • Inputs and results are deleted 24 hours after last use and are never used for training. 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.

Status

Free beta. Service page: https://render.janction.jp (/v1/health, /llms.txt, /faq). Blender 5.0, Cycles on GPU (blender="5.2" selects Blender 5.2 when the worker has it; render_info lists what is available). Guides: render farm for agents · without a GPU · Claude · ChatGPT · Claude Code · Codex · bpy scripts · API · comparison · Blender add-on · Maya / Houdini / Cinema 4D · 日本語は /ja/

Connect (remote MCP, nothing to install)

MCP server URL: https://render.janction.jp/mcp (Streamable HTTP). Auth is OAuth 2.1 with dynamic client registration: pressing "Connect" opens a page that creates a free API key (or takes one you already have). A raw API key also works as Authorization: Bearer jr_....

clienthow
Claude.ai (web, desktop, mobile)Settings → Connectors → Add custom connector → paste the URL → Connect
ChatGPTSettings → Connectors → Advanced → Developer mode → Create → paste the URL (OAuth)
Claude Codeclaude mcp add --transport http janction-render https://render.janction.jp/mcp, then /mcp to authenticate
Cursor, Windsurf, other MCP clientsStreamable HTTP at the URL above (OAuth, or a Bearer API key header)

Remote tools take scene_script (bpy code as text), scene_url (an https link to a .blend, .py or a 3D file) 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

Install (stdio MCP, sends local files)

The package is on PyPI as janction-render.

# 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.

Then, in Claude Code:

Build a small street scene in Blender with a camera fly-through and render a preview with janction-render.

Tools

toolwhat
`scene_info(scene_scriptscene_url
render_preview(..., frames="1-24", environment?, blender?)up to 4 frames (720p budget) tiled with frame labels; returns the image inline
render_estimate(scene_id, frame_start, frame_end, width, height, samples)GPU seconds, queue wait, "about N minutes", fits today's free quota? No GPU time used
render_final(scene_id, frame_start, frame_end, width, height, samples, fps, output, environment?, blender?, engine?, transparent?, notify_url?)frames (png / exr) or video (mp4 / webm / prores / gif / webp); transparent=True keeps an alpha background (png / exr / webm / gif / webp); notify_url gets one JSON POST when the job finishes; returns job_id + estimate
render_status(job_id)progress and ETA (eta.human); render_download(job_id, only="mp4" / "frames" / "all") files or links; render_cancel(job_id)
billing()free-beta quota (used today, daily limit, reset time); later balance and top-up link
render_share(job_id, title?, note?, include_script?, listed?)a public page /r/<id> with the image or video, the conditions and (optionally) the script; survives the 24-hour expiry until render_unshare; with listed=True it appears in /gallery after a review
asset_search(query, kind)CC0 models / textures / HDRIs on Poly Haven by words; results carry polyhaven:<id> and the entry file
render_info()workers online or gated (GPU lent to another workload), queue, expected wait, supported inputs, environment presets, Blender versions

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|info.

Writing a scene script

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)

The service sets the engine (Cycles by default; engine="eevee" for cheaper animations: about half the per-frame cost after ~7 s of shader compilation per job), resolution, samples and denoising; your script sets the scene, camera and frame range. 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. See samples/cube_scene.py.

HTTP 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
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}
GET  /v1/jobs/{id}      status, progress, eta, artifacts[], cost, warnings, info    DELETE /v1/jobs/{id}  cancel
GET  /v1/jobs/{id}/artifacts/{name}             PNG / MP4
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
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 /v1/me
GET  /llms.txt  /llms-full.txt  /ja/llms.txt  /faq  /terms  /privacy  /legal  /security  /support

During the free beta a 429 quota_exceeded response carries resets_at; a 400 beta_limit means the job is too big (split it). Once paid plans start, a 402 payment_required response carries checkout_url.

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-0.1.0.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

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.