Skip to content

Quickstart

From nothing to your first AI-built network in about five minutes.

The fastest way to experience Embody is the AI way: install one thing, drag in one file, click one button, then talk to TouchDesigner and watch it build. You describe what you want in plain language — no coding, no scripting, and git is entirely optional.

Embody itself is free, open source (MIT), and runs entirely on your own machine — no Embody account, no subscription, nothing SaaSy. You'll sign in to your AI assistant (a Claude account for Claude Code, for example), but that's the only login involved. The server Embody starts is local-only; nothing about your project leaves your computer unless you choose to share it.

Step 1 — Get the two prerequisites

You need two things installed before you start:

Never used an AI coding tool?

Claude Code is the gentlest place to start — it runs as a desktop app, a VS Code extension, or a web app, not only in a terminal. The first time, you'll create or sign in to a Claude account and then open a folder; you won't be writing code, just describing what you want in plain language. (The five-minute estimate above assumes TouchDesigner and your AI assistant are already installed — first-time installs take a little longer.)

Step 2 — Download Embody

Grab the latest Embody .tox from GitHub Releases. It's a single file.

Step 3 — Drag it into your project

Open your .toe project (or a brand-new one) and drag the .tox into the network. That's the entire install — there's nothing else to set up.

Embody initializes itself automatically over the next couple of frames. The core externalization features are self-contained and need no external dependencies.

Step 4 — Say yes to Envoy, the AI bridge

As soon as Embody finishes initializing, it asks whether you want to set up Envoy — its AI bridge. A dialog appears asking "Enable Envoy?" with two buttons, Skip and Enable Envoy. Click Enable Envoy.

That single click does everything for you:

  • Installs the MCP server's Python dependencies (~30 MB — TouchDesigner goes unresponsive for a few seconds while this runs)
  • Starts a local server on localhost:9870
  • Writes the AI config files into your project root — CLAUDE.md, AGENTS.md, .mcp.json, and a .claude/ folder of rules and skills — so your assistant knows how to talk to TouchDesigner

Clicked Skip, or want to turn Envoy off later? Toggle the Envoy Enable parameter on the Embody component anytime.

If it asks about git

Don't use version control? After you click Enable Envoy, you may see a second dialog recommending a git repo. Among its options, click Start Without Git (avoid Cancel — that stops the setup). The AI config files are generated either way, and everything works the same.

Local and private by default

Envoy only listens on your own computer (127.0.0.1) — it's not reachable from the internet or your network. Every Envoy tool is pre-authorized, too, so you're not stuck clicking through permission prompts.

Step 5 — Open your AI assistant and start talking

Keep TouchDesigner open — your AI assistant talks to the live session.

The easiest way in is the Launch AI Client button: on the Embody component's Envoy parameter page, set the AI Client menu to your assistant, then click Launch AI Client. Embody opens the client (an editor as a workspace, or a CLI in a new terminal) already pointed at your project root — and if the client isn't installed yet, it prompts you with instructions to install it first.

Prefer to open it yourself? Open your AI assistant in the same folder as your .toe and start a new chat:

  • Claude Code (the fully auto-configured path): open that folder — File → Open Folder in the desktop or VS Code app, or cd into it and run claude in a terminal — then start a session. It detects the .mcp.json Envoy generated and connects on its own.
  • Cursor or Windsurf: open the same folder; you may need to point it at the generated .mcp.json yourself — see Envoy Setup.

Now just say what you want. Try this:

Build me a noise-driven particle system.

Watch the operators appear in your live session — wired up, named, annotated, and laid out. Then keep going, conversationally:

Now make it react to audio, slow it down, and add a bloom on the final output.

You're not getting a screenshot or a code snippet to paste. You're getting the actual network, in front of you, ready to play with.

How long does a build take?

Small things — a parameter change, a fix, a question about your network — can land in seconds. A complete network is a real build: the AI reads your project, plans, creates, wires, and verifies its work, which typically takes 5-20 minutes of autonomous effort, depending on your model and compute. You don't have to watch it, and you don't have to run just one: Envoy coordinates multiple AI sessions on the same project, each scoped to its own part of the network. Less like waiting on a genie, more like directing a team.

Confirm the connection

If your assistant doesn't seem to see TouchDesigner, ask it to "list all operators in the project." If that comes back empty or errors, check that Envoy Enable is on and that you started the AI session after enabling it. Still stuck? See Envoy Troubleshooting.

What you got for free

Embody's other job is making your network version-controllable: any COMP or DAT can be saved to disk as readable text — something you can diff, review, restore, and hand back to the AI later. No binary black box, no lock-in.

Externalization is opt-in — nothing is written to disk until you choose. Two ways to opt in:

  • Tag operators yourself: select one and press Left Ctrl twice.
  • Auto-externalize what the AI builds: set the Auto-Externalize New Ops parameter (Embody component → Envoy page) to DATs, COMPs, or DATs and COMPs. New operators your assistant creates through Envoy are then tagged and externalized automatically. The default is Neither.

To save your changes, press Ctrl+Shift+U. On the next project open, everything tagged restores from disk automatically.

That's the whole loop — generate, compare, revert, branch. The lateral moves run at the speed of typing; the generating you delegate and come back to.

Here for version control, not AI?

You can skip Envoy entirely and use Embody as a pure externalization engine — every operator diffable on disk, no AI involved. See Getting Started for that workflow.

Where to go next

If you want to… Go here
Understand externalization in depth Getting Started
See everything the AI can do Envoy Tools Reference
Configure ports, multiple instances, or permissions Envoy Setup
Fix a connection problem Envoy Troubleshooting
Understand why this exists The Manifesto

The tool keeps up with you, instead of the other way around. That's the whole idea.