I would like to share with you my first properly working version of a tool I am working on. It’s an MCP Server designed to access and serve up to date TouchDesigner information straight from the official documentation.
The way it works is you’ll need VS Code (or similar programs like VS Codium) and the Roo Code extension. From there you’ll need to configure the Roo Code extension to use your favourite LLM. You can get API keys for free from Google Gemini and OpenAI I think but I use Claude API and I pay for my tokens.
I know this all sounds a little convoluted and I’m looking at taking this concept and creating a plugin or extension that will natively work inside of TouchDesigner but for now I can confirm it is really working.
I would appreciate any feedback you may have and I will do my best to answer any questions. I will say this though, I am a massive coding noob and still quite new to TouchDesigner myself though I am learning a lot through this project. Anyways I hope some of you find this useful or interesting enough to check it out!
We really appreciate you sharing this with the community. Your documentation server is a clever approach and tackling this as a learning project makes it even more impressive…
Also wanted to let you know that we have been experimenting internally with MCP integration for TouchDesigner. We’re still forumulating our appoach and havent settled on a direction yet, but didnt want to surprise you or others working on similar projects if we eventually release something official.
On our end we have established a real-time connection, with ability to walk the heirarchy, create nodes etc. As well everything is pure Python and REST API based. Its pretty cool but Claude really does not understand TouchDesigner. This is the current barrier to moving forward with things. Your angle to focus on pure documentation is astute.
Please keep experimeting and we are interested to see where this project leads you.
Oooooh! Even the possibility of an official MCP for Touchdesigner is exciting - fingers crossed.
I don’t fully understand them just yet, but as with TD, learning is part of the fun for me…down the rabbit hole I go
Ooo both excited to check out the community tool and what Derivative has in mind for this. I often can get lost in the sauce when things start to get buried in containers or a spaghetti of connection lines. Documentation, organizing, data wrangling and manipulation would be so much more powerful for me with MCP servers. Been using Cursor for some projects and previously used earlier versions of claude (before Claude code, pre-sonnet 3 - i think) to draft R/python functions to create datasets usable in TD.
Hi! Just wondering if there is any way to currently port this directly with CLI instead of having to use VS Code. This is an amazing service and I would love to be able to use it directly from the CLI instead.
Hey just a quick follow-up! I forked your repository and make a version that is completely compatible with Codex! Let me know if you’re interested in adding that as a contribution
We have something internal that is working well but we are still testing and contemplating what to do next still… In the meantime this looks interesting…
really cool to see this. looks super comprehensive which is perfect for a project like this.
i’ve been working on something related: an mcp server that runs inside touch, so ai assistants can actually create/modify operators, understand network layouts, connect ops, set pars, execute python, etc… it’s incredible to me how well it works out-of-the-box, but there’s a long journey ahead for sure…
I have also been developing an MCP server using an OpenAPI-compliant Python API server since April 2025, and have released it as open-source software as the most popular TouchDesigner MCP server.
I hope the official developers will release a reliable MCP server soon. Good luck with your project!
This is really exciting to see. Thanks for sharing it, and also thanks to everyone else in this thread who has been experimenting around TouchDesigner + MCP.
I have been working on a related open-source project called tdmcp / MindDesigner:
My approach combines both sides that people have mentioned here:
an embedded TouchDesigner knowledge base, so the assistant has real operator/context awareness
a live bridge inside TouchDesigner, so the assistant can create, modify, inspect, connect, arrange, and preview real networks
The current version exposes 179 MCP tools, from high-level generators like feedback networks, audio-reactive systems, particles, Shader Park scenes, and VJ/live tools, down to lower-level node CRUD, parameter editing, graph inspection, project analysis, reusable components, recipes, and vault/documentation workflows.
There is also a Claude Desktop .mcpb bundle for a no-terminal install:
I am especially interested in the same question Jarrett mentioned earlier in the thread: the assistant needs more than just access to TouchDesigner; it needs enough TouchDesigner-specific understanding to make sensible networks. That has been the main design goal of tdmcp: combine real execution with enough TD knowledge that the assistant is not just guessing operator names.
Would love feedback from anyone testing MCP workflows in TouchDesigner, especially around real creative workflows, live-performance safety, useful high-level tools, and places where AI assistants still misunderstand TD concepts.
hey Dylan, loved reading The Great Inversion ! went through it a few weeks ago and keep coming back to the ideas. one of the best TouchDesigner think-pieces I’ve come across (would def love to find more of these thought-provoking reads on TD and multimedia art creation in general) and have started using Embody as well to some pretty interesting results.
This is really interesting, thanks for sharing it.
I’ve been looking into how TouchDesigner users move between documentation, AI tools, and real project workflows, especially when a project involves external inputs or multiple tools. A TouchDesigner-focused MCP server feels useful for beginners who constantly jump between the docs, examples, and their network while learning.
One question: does it mainly retrieve documentation, or can it also help an LLM reason about operator parameters and suggest small example networks?
Curious to see how people start using MCP around TouchDesigner.
I’ve recently released Embody v6. It’s more than just an MCP server – it parses and writes to TDN, an open network file format that’s YAML-based.
What’s great about TDN is that it makes small example networks far more portable and sharable now. I’ve added a community feature called “the collection”:
This lets users submit their own “specimens” (sharable COMPs) – no download necessary, simply click the copy button or copy the TDN text to your clipboard, and TouchDesigner running with Embody v6+ will automatically pick them up.
The TDN format is 100% open source – the goal here is that the more networks we share and publish, the more models can be trained on them to improve the network creation quality.
On top of all of this, Embody comes bundled with quite a few Claude skills and rules that help it understand the context of Touch.
