Claude Opus 5.5 used MCP to build a VFX scene in Blender and Nuke without a single AI-generated frame

Almost every video about “AI in filmmaking” follows the same formula: write a prompt and get beautiful footage you cannot control. The director cannot say, “Move that lantern to the left and make the leaves heavier.” This story takes a different approach. A creator from the world of film and post-production did not generate any images at all. Instead, they put Claude Opus 5.5 in charge of a conventional VFX pipeline.

What they actually built

According to the post, Opus 5.5 was connected to Blender and Nuke through MCP (Model Context Protocol). Houdini was also used. Rather than relying on an off-the-shelf solution for the wind, lanterns and leaves, the model wrote its own physics engine in Python and NumPy.

Claude then worked as a VFX supervisor. It determined the composition, silhouettes, architecture, object movement, lighting and fine details of the animation. The human provided feedback, and the model translated it into changes to the scene. According to the creator, there were more than 580 rounds of feedback and revisions, and they never once edited the scene by hand. Everything was controlled through conversation.

How this differs from video generation

Generative video delivers finished pixels. Changing a single detail without risking damage to everything else is almost impossible. Here, the model produces a scene: geometry, materials, lighting, animation, a compositing graph and simulation code. Any element can be opened, changed and rendered again. It is the familiar studio iteration cycle of “revise—render—review,” with the model doing the execution.

MCP also plays an important role. It is a standard way to give a model access to external applications. Blender and Nuke become sets of tools it can use: read the current state, perform an action, retrieve the result. Previously, this kind of integration was more often demonstrated on simple tasks. Here, according to the creator’s account, it handled hundreds of successive revisions to a single scene.

The fact that the model wrote the physics engine itself is particularly interesting. A familiar situation: none of the presets can produce the effect you need, which used to mean handing the task to a technical artist or programmer. Now, that piece of code can be created in the same conversation used to discuss the shot.

What remains out of view

There is only one original source here: a social media post, recapped in a Telegram channel. Much remains unknown:

• how much time and how many tokens the 580 revisions consumed;
• how the model “saw” the result: whether it examined renders or worked from descriptions and numbers;
• how often unsuccessful steps had to be rolled back;
• how well the final video would meet production requirements, beyond serving as a demonstration.

The claim of “no manual intervention” also needs to be treated carefully: the creator may have set the tasks and approved the results, which already accounts for a substantial share of directing. Their background in film and post-production was probably the key resource here. The human knows what feedback to give and how to express it in the language of the craft.

What this means in practice

For developers and product managers, the takeaway is straightforward: a useful application for agents is controlling professional tools through MCP or APIs, rather than generating the final output. This preserves control and keeps the result editable.

If you are building a product for creative professionals, consider this pattern: conversation as the interface, and a project a human can understand as the output. For studios, the implication is different: a skilled practitioner paired with a capable model could handle work that previously required a small team. But that proposition should be tested on your own projects, rather than judged solely by an impressive video.