AI agents in creative tools: MCP opens the door

The biggest practical news from this year's SIGGRAPH is the massive arrival of the Model Context Protocol (MCP) in creative applications. MCP is an open standard that allows AI agents to communicate with applications through a standardized interface — similar to USB for hardware, but for software and AI. And at SIGGRAPH, virtually the entire top tier of the creative industry adopted it.

Adobe is expanding its creative agent across Firefly, Express, and Creative Cloud. The result is an assistant you can tell what you want to create, and it will organize a multi-step workflow on its own. Adobe is also releasing the Adobe Express Developer MCP Server, which will let developers build add-ons for Adobe tools with the help of AI assistants.

Blender — the free open-source 3D tool that is exceptionally popular among Czech graphic artists and studios — now offers an MCP server via Blender Lab. This means that an AI agent can directly control Blender through its Python API: creating objects, manipulating the scene, or helping with technical tasks that normally take hours.

Epic Games announced MCP connectivity for Unreal Editor, which will let game developers and virtual production teams use AI assistants that understand the scene, assets, and project state. SideFX is integrating MCP into Houdini 22 for procedural character rigging. Affinity by Canva connected its tools with Claude via MCP, so designers can automate batch edits, layer renaming, or export preparation.

What does this mean in practice? A graphic artist in Blender can ask an AI agent to check the scene for missing textures or prepare export variants. An editor in Adobe Premiere can let an agent prepare proxy files for daily reviews. And all without leaving the application itself.

Synthetic video detection: A tool for newsrooms and media

With the growing flood of AI-generated content, NVIDIA introduced the Synthetic Video Detector NIM microservice at SIGGRAPH — a tool that analyzes video frame by frame and determines the probability that the content was generated by artificial intelligence. Accuracy reaches 92% on uncompressed video, 87% at 15% compression, and still a solid 82% even at half compression.

For newsrooms, television stations, and news organizations — including Czech ones — this is a practical tool that can be deployed in editorial workflow. Processing 1080p video takes just 22 milliseconds on NVIDIA RTX systems. Partner Wowza is already integrating the detector into its Video Intelligence Framework, which is used by tens of thousands of deployments in more than 170 countries. The detector can also be operated in air-gapped environments without internet access, which is crucial for security-sensitive institutions.

Cosmos 3 Edge: Physical AI at the network edge

NVIDIA released Cosmos 3 Edge, a lightweight 4-billion-parameter model from the Cosmos series, optimized to run directly on devices — from Jetson Thor to RTX PRO to regular GeForce RTX GPUs. The model understands and generates text, images, video, audio, and robot actions and is freely available on Hugging Face under an open license.

Cosmos 3 Edge achieves first place in the VANTAGE-Bench benchmark in its performance class. For robotics, this means a robot can run an advanced world model locally, without relying on a cloud connection. Partners such as Agile Robots, Doosan Robotics, and Siemens are already testing the model for industrial deployment.

For autonomous vehicles, Cosmos 3 Edge offers traffic scene understanding, object intention prediction, and control model distillation — all on hardware with limited resources. The platform is now available in three sizes: Edge (4B), Nano (16B), and Super (64B), so developers can choose the right model for each stage of the project.

DGX Station and NemoClaw: A super-agent on your desk

At SIGGRAPH, NVIDIA presented the DGX Station in detail, featuring the GB300 Grace Blackwell Ultra Desktop Superchip — a desktop supercomputer with 20 petaflops of FP4 AI performance and 748 GB of coherent memory. Combined with the NVIDIA Agent Toolkit and the open-source NemoClaw blueprint, the system can be up and running in about 30 minutes and run full-fledged AI agents — completely offline.

The key model is Nemotron 3 Ultra with 550 billion parameters, running as an open model directly on the DGX Station. LangChain optimized its Deep Agents harness on it, and Nous Research fine-tuned the model for its production agents. For companies, this means a fundamental economic change: after a one-time hardware purchase, there are zero per-token inference fees.

The DGX Station can be ordered from manufacturers like ASUS, Dell, HP, MSI, GIGABYTE, and Supermicro — brands that are commonly available on the Czech market as well. Pricing has not been officially announced yet, but given last year's models, it will be in the range of higher single-digit millions of crowns.

MotionBricks and 21 research papers: From animation to physical robots

The research portion of SIGGRAPH brought 21 technical papers from NVIDIA, united by one theme: virtual worlds must not only look realistic but also behave realistically. The most visible demonstration is MotionBricks — a real-time motion model trained on over 350,000 motion clips. The same model that animates a character on screen simultaneously controls a physical humanoid robot, the Unitree G1.

Connecting virtual animation and physical robotics is precisely the direction NVIDIA pursued at SIGGRAPH. The GPC framework creates the foundation for a "foundation model of motor control" — a model that learns general movement skills and transfers them to new tasks. ArtiFixer, on the other hand, converts raw 3D scans of the real world into clean, complete virtual scenes.

What it means for Czech creators and businesses

For the Czech creative scene, the most exciting development is the adoption of MCP in Blender — the open-source tool widely used in the Czech Republic, from independent creators to professional studios. The ability to have an AI agent working directly inside Blender lowers the barrier of entry for smaller teams and individuals.

Cosmos 3 Edge on Hugging Face means that anyone with a GeForce RTX card can experiment with advanced world models. The Synthetic Video Detector will be relevant for Czech media and newsrooms, which are increasingly facing the question of whether the video they received is real. And the DGX Station — although unaffordable for most — shows the direction development is heading: AI not only in the cloud, but also locally on the desk.