Blender 5.3 reaches JANCTION Render within days, drivable by AI agents over MCP

Blender 5.3 shipped this week, and the cloud rendering service JANCTION Render is already running it on remote GPUs. What makes the pairing interesting is the interface: an MCP endpoint that lets an AI agent write Blender Python scripts, submit them, and receive finished images — with no local copy of Blender and no local GPU. A developer walking through the setup on dev.to documented the whole loop, including two mistakes the agent made and the measured time and cost of every step.

How the setup works

According to the dev.to post, the author pointed an agent at JANCTION Render's MCP server. The agent writes a bpy script, the service executes it in Blender 5.3 on its own GPUs, and images come back. Each preview reportedly returns a short automated assessment — labelled ok, check or fix — together with a quote for the full render before any charge is made. The service also offers a browser route for uploading an existing .blend file, and the author says the MCP connection works with agents in Claude, Cursor or Codex.

The test scene was chosen to exercise what 5.3 can do: three faceted glass gems that split light into spectral colour. The author describes asking for it in a single sentence and letting the agent write its own roughly 90-line script — a grey floor, emissive strips at the back, one spotlight, three icospheres carrying the glass material, and a camera.

Where the agent went wrong

The first preview returned in five seconds and the gems were nearly black, with no colour at all. The post gives two causes. The script looked for a Principled BSDF input named "Dispersion", but in Blender 5.3 the dispersion controls sit under Transmission, with inputs such as "Transmission Weight", "IOR", "Transmission Dispersion Scale" and "Transmission Dispersion Abbe Number". The missing input was silently skipped rather than raising an error. On top of that, the dark studio gave the glass nothing to work with — glass renders its surroundings, and there were effectively none.

The fix was a light grey floor, one narrow spotlight behind and above the gems, and the corrected input names. The next preview showed colour separation along the facets and passed the automated check.

The second mistake surfaced in the animation. The author wanted the gems to rotate across 72 frames; a ten-frame trial looked fine, but the full job came back as 72 copies of the same picture. The cause was ordering: the script set the rotation value and then called frame_set(), which re-evaluates the animation and discards the value just written. Passing the frame number directly to keyframe_insert instead fixed it — and, as the post notes, rendering a short contiguous test batch before the full job is what caught the bug cheaply.

Time and cost

The author's measurements, taken on 2026-11-11:

• Preview, 1280x720, 32 samples, one frame: 2.8 GPU-seconds, 5 seconds wall time, 1 JPY.
• Ten-frame test, 1920x1080, 128 samples: 59.9 GPU-seconds, 19 seconds, 6 JPY.
• Final job, 1920x1080, 128 samples, 72 frames: 267 GPU-seconds, 1 minute 25 seconds, 27 JPY.

That averages about 3.7 GPU-seconds per 1080p frame. The final job ran in parallel chunks, so its wall time came in well under the total GPU time. Per the post, new keys get some free GPU time, after which the service charges 0.1 JPY per GPU-second. These figures come from a single author's walkthrough, not an independent benchmark.

Why it matters

The lag between a major 3D application release and its arrival in remote rendering workflows has shrunk to days. A feature like glass dispersion, new in 5.3, was exercisable from a chat window in the same week it shipped, with zero local setup — which is exactly what cloud rendering with an agent-friendly protocol is supposed to deliver.

The workflow itself is the other story. Cheap previews, an automated quality check, and a small paid test batch before committing to the full job form a sensible template for agentic use of expensive compute: the animation bug cost 6 JPY to find instead of surfacing after a finished render. And the silent failure mode — an agent guessing an input name that no longer exists — is a reminder that version drift in scripted APIs breaks quietly, which is precisely where a render-and-inspect loop earns its keep.