AI Is Rewriting Creative Coding: How Artists Use Claude, TouchDesigner, and Code to Make Generative Art

Scroll short-form video for five minutes and you'll hit one: a glowing particle cloud that pulses to the bass, a face dissolving into flowing noise, a webcam feed melting into a painting in real time. A huge share of those clips come out of one tool, TouchDesigner, and more and more of them are now built with AI rather than just by AI.

That distinction matters. This isn't "type a prompt, get a picture." It's artists using AI as a creative coding partner: it writes the shader, wires up the node network, debugs the Python, and leaves the human in charge of the idea, the taste, and the final call. The art is still code. The AI just removed most of the friction between "I wonder if…" and seeing it on screen.

This guide breaks down what's happening, how the TouchDesigner workflow actually works, and how to start, even if you've never written a line of code.

Two kinds of "AI art," and why the difference matters

People lump everything together as AI art, but there are two very different practices:

Generative art has always been about building systems, not single images. Artists like Vera Molnár and Sol LeWitt were writing rules for art decades before GPUs. AI-assisted creative coding fits that tradition: you're still the author of the system. The AI is a very fast, very patient collaborator who knows the documentation by heart.

Why TouchDesigner is the center of the trend

TouchDesigner, made by Derivative, is a node-based visual programming environment for real-time graphics. You connect operators (noise generators, particle systems, audio analyzers, cameras, shaders) into a network, and the result renders live. It's been the backbone of concert visuals, museum installations, and projection mapping for years.

Three things pushed it into the mainstream feed:

1. It's free to start. There's a free Non-Commercial edition, so anyone can download it and follow a tutorial tonight. (Check Derivative's license terms before using it for paid work.)
2. The output is made for short video. Audio-reactive, loopable, hypnotic visuals are exactly what performs on TikTok, Reels, and Shorts. Freelancers now sell TouchDesigner visualizers for musicians' social content.
3. AI plugs in everywhere. TouchDesigner runs Python inside every project, speaks OSC and MIDI, and can host machine learning models. That makes it an ideal canvas for AI tools.

The catch has always been the learning curve. Hundreds of operators, Python expressions, GLSL shaders, and a lot of "why is this black?" debugging. That's exactly the part AI is eating.

How artists are using AI with TouchDesigner right now

1. Agents that build the network for you (MCP)

This is the newest and most exciting piece. MCP (Model Context Protocol) is an open standard that lets an AI agent use external tools. Derivative has released TDMCP, described as the official MCP server for TouchDesigner. It's in beta, requires TouchDesigner 2025.33070 or later, and ships setup presets for agents including Claude Code.

Once it's connected, you can tell an agent something like:

"Build an audio-reactive particle system. Drive particle size from the low frequencies of the incoming audio, color from the highs, and add a feedback trail. Expose speed and trail length as custom parameters."

The agent creates the operators, wires them, sets parameters, reads back errors, and fixes them. Community-built servers do similar things: create, modify, and delete operators, inspect parameters, and execute Python. Some even bundle a reference of every real operator so the AI doesn't invent ones that don't exist. There are also skill packs that teach an agent TouchDesigner conventions before it touches your project.

2. AI as a code writer inside the network

Even without MCP, the most common use is simple: ask Claude or another assistant to write the bits of code TouchDesigner needs.

• GLSL shaders. Describe the look ("domain-warped noise with a slow color cycle, palette from deep teal to hot pink") and get a pixel shader you paste into a GLSL TOP. Shaders are where AI saves the most time, because the math is unforgiving and the results are instantly visible.
• Python scripts. Automate repetitive setup: generate 50 instanced objects, build a preset system, map MIDI controller knobs to parameters.
• Expressions. Quick one-liners like animating a parameter over time with absTime.seconds.

Here's the flavor of Python an AI will write for you in TouchDesigner, creating a noise texture and running it through a Level TOP:

# Run inside TouchDesigner: build a tiny noise → level chain
parent_comp = op('/project1')
noise = parent_comp.create(noiseTOP, 'noise1')
level = parent_comp.create(levelTOP, 'level1')
noise.outputConnectors[0].connect(level)

It's tiny, but multiply it by every fiddly setup step in a big project and the hours add up fast.

3. Real-time diffusion: AI that paints live

The clips where a live camera feed turns into an oil painting or a dreamscape usually come from StreamDiffusion, a real-time diffusion pipeline. In TouchDesigner it's most often used through DotSimulate's StreamDiffusionTD component. It needs Windows and an NVIDIA GPU with CUDA, so it's a hardware commitment.

The creative move is feeding diffusion something generated, not just a camera: your particle system, your noise field, your audio-reactive geometry. TouchDesigner provides structure and motion; diffusion provides texture and style. Some artists go further and train a small style model (a LoRA) on their own work, or on a collaborator's paintings with permission, so the live output carries a genuinely personal look instead of a generic AI style.

4. AI as the performer's co-pilot

A more experimental pattern: an AI agent controls a running project in real time, sending OSC messages that tweak brightness, speed, or color based on a text description of a mood or the energy of a set. Tutorials pairing the Claude API with TouchDesigner over OSC are already out there. It's early, but it hints at visuals that respond to the room as well as the audio.

The workflow: from idea to clip

Here's the loop we see working best, whether you're a total beginner or a working VJ:

1. Start with a reference, not a prompt. Save three clips you love. Describe why they work: motion, color, rhythm, texture.
2. Ask the AI to explain before it builds. "What operators would create this effect, and how would they connect?" You learn the tool while you go.
3. Build in small steps. One effect at a time. Get the noise right, then the color, then the audio reaction. Big one-shot requests produce big tangled networks.
4. Expose the knobs. Ask the AI to surface the 3–5 parameters that matter as custom controls. This is where your taste comes in: tuning is the art.
5. Record, loop, post. Export a loop at your platform's aspect ratio. Vertical 9:16 for short-form.
6. Keep what you learned. Save working networks as reusable components (.tox files) and save prompts that worked. Your library becomes your style.

A starter prompt that works well with Claude:

I'm learning TouchDesigner (Non-Commercial, latest build).
Goal: a 9:16 audio-reactive visual for a 15-second loop.
Look: slow, organic noise; particles that swell on kick drums;
palette deep teal → magenta.
First explain the operator chain and why, in plain English.
Then give me step-by-step build instructions, one stage at a time.
Wait for me to confirm each stage works before the next.

Beyond TouchDesigner: the wider creative coding scene

TouchDesigner is the headline, but the same AI-plus-code pattern is everywhere:

• p5.js is the friendliest entry point: JavaScript in a browser, no install. The p5.js project even publishes its own tutorial on using generative AI to write p5.js code, including the critical questions artists should ask while doing it.
• Shaders (GLSL) on sites like Shadertoy, where AI-written fragment shaders are now a common starting point.
• Research tools are pushing further. One recent academic system lets people evolve shader art just by picking outputs they like, with an LLM rewriting the shader code behind the scenes, so no coding is needed at all.
• Claude Code skills for p5.js, GLSL, and generative art are circulating in community directories. They're unofficial, so read them before installing.

Artists and builders to watch

The people below don't all use TouchDesigner or chat assistants, but each one shows where AI-plus-code art is heading. Study them for ideas, not shortcuts.

• Refik Anadol turned AI into museum-scale spectacle. His studio trained a model on MoMA's digitized collection for Unsupervised (MoMA, 2022–23) and explored its latent space with custom software on a roughly 24-by-24-foot lobby wall. It's the template for the "data sculpture" look you see imitated everywhere.
• Sougwen Chung draws and paints live alongside robots, the DOUG series, guided by AI. Her 2025 installation Spectral combined robotic painting with AI trained on her earlier work and real-time EEG input. It's the clearest example of AI as a performance partner rather than a generator.
• Holly Herndon and Mat Dryhurst built The Call (Serpentine, 2024–25) around vocal datasets made with choirs, and released Holly+, an AI model of Herndon's voice others can use. They're the leading voices on consent and data ownership in AI art, which matters if you plan to train your own models.
• DotSimulate builds StreamDiffusionTD, the component behind much of the real-time diffusion work in TouchDesigner. Following the tool's development is the fastest way to see what's newly possible live.
• The Derivative community publishes tutorials, project files, and artist posts on derivative.ca, including a growing number on StreamDiffusion and MCP. It's the best place to find working TouchDesigner artists who share their process.

The honest part: credit, originality, and skill

Hype is fun, but a few things are worth saying plainly:

• The AI didn't have the idea. The best work in this space has a clear point of view that no prompt produced. Use AI to go faster, not to skip having taste.
• Learn what the code does. If you can't read the shader, you can't push it somewhere new. Ask the AI to explain every block it writes. That's how this becomes a skill and not a dependency.
• Respect other artists. Training a style model on someone's work without permission isn't a gray area. Use your own work, licensed material, or collaborators who said yes.
• Disclose your process. Audiences increasingly care how a piece was made. "Built in TouchDesigner, code co-written with Claude" is a flex, not a confession.

Where to start this week

1. Download TouchDesigner (the free Non-Commercial edition) and follow one beginner audio-reactive tutorial end to end, without AI, so you understand the basics.
2. Rebuild it with an AI. Ask Claude to explain each step, then to extend it with one new effect.
3. Try MCP once you're comfortable. Connect Claude Code to TouchDesigner through TDMCP on a copy of your project and let the agent do the wiring.
4. Level up the fundamentals. If you want to understand how agents, tools, and MCP work under the hood, our review of Building with the Claude API covers the course that teaches exactly that. For free foundations, see the Anthropic Academy picks in our Claude fall update.

The bottom line

Creative coding used to have a brutal on-ramp: months of fighting syntax before you made anything you were proud of. AI has flattened that ramp. Agents can now operate TouchDesigner directly, assistants write shaders on demand, and real-time diffusion turns any generated motion into a living painting.

What hasn't changed is that the art lives in the choices: what to build, what to keep, what to throw away. The tools just got out of the way. If you've ever wanted to make the kind of visuals that stop people mid-scroll, there has never been a better time to start.

More AI workflows live in AI & Workflows. If you're using Claude for technical work beyond art, see Claude Code for cybersecurity workflows.