Chaos released V-Ray 7, Update 5 on 5 November, and the part of it that matters to anyone tracking agents in design software is not the thirty-odd new features. It is that the company has published Model Context Protocol servers for three of its products, and that one of them lets an outside AI assistant change things in a live scene.
The release post on Chaos’s blog, dated 5 November, covers the full update: new lights and volumetrics, an AMD GPU rendering path through HIP, a node-based material editor for the SketchUp and Rhino plugins, Spanish localisation for V-Ray for SketchUp, and 240 new animated 4D characters in Anima. AEC Magazine’s write-up, posted the following day, counts more than thirty new features across V-Ray, Chaos Vantage, Chaos Veras and Chaos Anima, all reachable “through a single V-Ray Collection licence.”
The MCP work is the structural change.
What the assistant is allowed to touch
Chaos’s MCP servers page describes the servers as “a bridge between compatible AI assistants and Chaos visualization products,” and names ChatGPT and Claude as examples of compatible assistants. Three products have servers at launch:
- V-Ray. The assistant can inspect scene information, adjust supported lighting and material parameters, manage cameras, and modify rendering and environment settings.
- Chaos Cosmos. It can search the asset library for models and materials and place supported assets in a scene.
- Chaos Anima. It can simplify repetitive operations and populate a visualisation with people and crowds.
Read that first bullet again. Inspecting a scene is a read. Adjusting lighting and material parameters, managing cameras and changing render settings are writes — the assistant is moving the controls, not describing them. This is the same line we traced through Autodesk’s three Revit MCP servers, where the distinction between a server that reads a model and one that edits it turned out to be the whole governance question. Chaos’s V-Ray server is on the write side of it, with the important qualifier that what it writes is the look of a scene rather than its geometry. A bad render is a bad render; it is not a wrong dimension handed to a fabricator.
What the page does not do is enumerate the limits. Every capability is scoped with the word “supported” — supported lighting and material parameters, supported assets — and no complete list of what falls inside that word appears on the page. Chaos’s own guidance is that users should inspect and refine results with the professional controls in their software, which is the honest framing for a tool at this stage, and also an admission that the agent’s output is a draft.
The host list is the real constraint
Here is where the announcement is narrower than it reads. The V-Ray MCP server is available for SketchUp and Rhino. For 3ds Max, Chaos says a dedicated server is not needed, because third-party 3ds Max MCP servers can already reach V-Ray objects, lights, materials, render settings and the V-Ray Frame Buffer — a reasonable answer, but one that puts a 3ds Max user’s agent workflow on software Chaos does not ship. An MCP server for Chaos Vantage, the real-time viewport, is listed as coming soon, with no date on either the MCP page or the release post.
There is also a version floor: the page tells users to “Access the new MCP capabilities with V-Ray 7, Update 5.” Agents are not arriving as a cloud feature that appears in an old install.
So the set of people who can point Claude or ChatGPT at a V-Ray scene today is: those on Update 5, in SketchUp or Rhino, plus 3ds Max users willing to run a third-party server. Those are widely used plugins, so that is not a small population — but it is not “V-Ray now has MCP.”
The assistant that ships inside the product
Separately from the servers, V-Ray now has its own AI Assistant built into the interface, available in V-Ray for SketchUp, Rhino and 3ds Max. Per the release post it offers guidance as you work, helps explore lighting, and finds assets; it includes Sky Match, which matches a scene’s lighting and mood to a reference image, and surfaces asset recommendations from Cosmos. Chaos’s What’s New page also lists Render Critique under the AI Assistant — feedback on a finished render — which AEC Magazine mentions and the release post does not.
The division of labour is worth noting. The built-in assistant covers the three hosts; the MCP server covers two. If you want an agent you already pay for, running your own prompts and your own context, you are on the shorter list.
Cosmos gets two additions of its own: image-based search, where a reference image returns visually similar models and materials, and AI 3D generation, which turns a single 2D reference image into a 3D model. Anima gets AI pathfinding, letting characters navigate a scene on their own while accounting for slopes, obstacles and crowd flow, plus a Waypoint tool for steering them.
Chaos names the models
The detail most vendors leave out is in the Veras section. Veras 5.3 adds a NODES tab for building node-based workflows, available only in the standalone Veras app, and new video models — Veo 3.1, Grok Imagine and Gemini Omni, with output up to 4K and native audio — alongside updated OpenAI image models. Its new Agent mode builds 3D models from up to twelve reference images, and the release post says it runs on GPT-6 Astra or Claude Opus 5.5. Agent mode, like the NODES tab, is currently available only in the standalone Veras app — another capability whose reach is narrower than the headline.
That is a vendor telling customers, in writing, which frontier models sit under a feature and letting them choose between two. It is the same instinct we saw when Qonic shipped per-request model choice and token-metered AI inside its BIM platform: disclosure of the engine, rather than a generic “powered by AI.” The reason it matters commercially is that frontier-model pricing moves, and a studio that knows which model it is calling can reason about what a heavy week costs. Chaos’s release post does not say how Veras Agent usage is metered, and the What’s New page does not state pricing for the update; neither document answers what a studio should budget.
Where this sits
The MCP wave in AEC software has moved fast enough this autumn to have a shape. Nemetschek turned “Max” into a tier across Archicad, Allplan and Bluebeam with MCP servers attached. Autom8Labs gave away 22 MCP skills on the argument that the connector was never the hard part. Qonic put agents inside the model. What distinguishes the Chaos release is where in the stack it lands: the renderer, the asset library and the crowd tool, rather than the authoring environment where most of this autumn’s MCP work has gone.
Analysis: the interesting question is whether visualisation is an easier place for agents to earn trust than modelling, and there is a case that it is. A rendering is reviewed by eye before anyone acts on it. A model edit propagates into schedules, quantities and shop drawings whether or not anyone looked. If agents are going to learn the AEC domain in public, the stack where the failure mode is an ugly picture is a gentler classroom than the one where the failure mode is a wrong quantity.
Allan Poore, Chaos’s chief product officer, framed the update to AEC Magazine this way: “Every industry is going through a push and pull with AI right now, as they attempt to figure out what’s actually useful.” The update, he said, “justifies itself on day one, not only by dramatically shrinking the V-Ray learning curve, but also allowing creators to direct the process across assets, automation, and analysis, making their visions much easier to attain.”
The learning-curve claim is the one to watch. V-Ray’s settings panel has long been a specialist skill, and an assistant that can answer “why is this render noisy” without a forum search is a real reduction in the cost of competence. Whether it also reduces the cost of a good image is a different question, and not one a release post can settle.
