Skip to content

Embody

Create at the speed of thought

Embody puts a team of AI builders inside your live TouchDesigner project. Describe a network in plain language and it gets built — wired, named, annotated, and laid out while you direct. Run several agents in parallel, each scoped to its own part of the network. Branch off the version that works, restore yesterday's state. The tool keeps up with you, instead of the other way around.

Quick Start

  1. Download the Embody .tox from the release folder
  2. Drag and drop it into your TouchDesigner project
  3. Enable Envoy to connect AI assistants to your session
  4. Tag operators by pressing Left Ctrl twice on any COMP or DAT
  5. Save with Ctrl+Shift+U — on next project open, everything restores from disk automatically

Five-minute Quickstart Full setup guide

Embody is four tools working together — lateral velocity, forward velocity, outward velocity, and the substrate that makes it all possible.

  • Embody — Lateral Velocity

    Tag any operator and Embody externalizes it to files on disk that mirror your network hierarchy. Try a new direction, branch off a good one, restore yesterday's state — all in seconds. Externalized files are the source of truth, so every project opens already in flow.

    Get started

  • Envoy — Forward Velocity

    An embedded MCP server with 68 tools that lets Claude Code, OpenCode, Codex, Gemini, Cursor, Windsurf, and GitHub Copilot via VS Code talk directly to your live TouchDesigner session. Say what you want — operators, connections, parameters, extensions, fixes — and watch it happen in your live session. Small changes can land in seconds, depending on your model; a complete network is a 5-20 minute autonomous build you can hand off, parallelize across sessions, and check back on.

    Setup Envoy

  • Convoy — Outward Velocity

    Convoy-enabled Embody nodes on a trusted LAN discover, inspect, and control each other — one AI session relaying builds, test runs, saves, screenshots, restarts, and Embody updates to every machine in the room. A small per-user background app keeps each node reachable even while TouchDesigner is closed.

    Join the Convoy

  • TDXN — The Substrate

    TouchDesigner networks exported as human-readable YAML. The format is what lets your AI agent see what's on screen, what lets you diff one attempt against another, and what lets a network rebuild itself from text. TDXN is what makes the rest of this possible.

    Learn about TDXN

What You Can Do

You and your AI assistant, in the same session, with full context of your live network:

Capability Example
Build entire networks from a sentence "Build me a noise-driven particle system."
Try a different approach without starting over "Actually, make it react to audio instead."
Read & set any parameter "Set the noise frequency to match the audio input."
Write extensions "Create an extension class that manages scene transitions."
Debug errors "Why is my render chain producing a black output?"
Compare attempts side by side "Show me what changed between this version and the last one."
Inspect anything "What parameters are non-default on this operator?"
:material-annotation: Document networks "Add annotations to group and label these operators."
Run tests "Run the test suite and fix any failures."

You describe what you want. The AI works with your live network — operators, connections, parameters, hierarchy — with the whole picture. And you don't have to run just one: multi-session coordination lets several assistants work the same project at once, each claiming its own scope, aware of the others, gated from collisions — an entire team on your network. The result is a network you can read, revert, and rebuild from text.

Full tool reference

Key Features

Feature Description
Automated Externalization Tag COMPs and DATs — or opt in to auto-tagging what the AI builds — and Embody keeps the external files in sync, auto-restoring from disk on project open
Envoy MCP Server 68 tools connect AI assistants to your live TD session
Multi-Session Coordination Run several AI sessions on one project — scoped claims, peer advisories, and gated destructive operations
Convoy LAN Relay Convoy-enabled nodes on a trusted LAN discover, inspect, and control each other — relay test runs, saves, screenshots, restarts, and fleet-wide Embody updates to other machines from one AI session (Convoy guide)
TDXN Format Export/import operator networks as diffable YAML for code review and snapshots
Keyboard Shortcuts Double-tap Left Ctrl to tag, Ctrl+Shift+U to save — and every binding is remappable (type or record)
Parameter Tracking Automatically detects parameter changes and marks COMPs dirty
145 Test Suites Comprehensive automated testing framework (4,551 tests)
Structured Logging Multi-destination logging with file rotation, ring buffer, and MCP access

Requirements

  • TouchDesigner 2025.33230 or later (Windows / macOS)
  • A git repository is optional. Embody works in any project folder; if you happen to use git, every change is also a clean diff for free.

Credits

Embody stands on other people's work, and this section names it. Ideas move between the TouchDesigner MCP projects in both directions; where Embody took one, it is credited here and in the code where it landed.

  • External Tox Saver by Tim Franklin -- Embody began in 2020 as a refactor of it.
  • TDMCP by Derivative (their experimental MCP server, a private repository at the time of writing, used with permission) -- undoable MCP mutations (one undo step per batch), the get_docs design (version-exact offline help first, the wiki API second), sequence-block growth on parameter writes, the transport-security hardening prompt, and treating a code-review rule as a query. Landed from v6.0.87 (2026-07-04) on.
  • TDMCPSkills by Derivative (private at the time of writing, used with permission) -- the pop-networks skill is adapted from td-pop-family.
  • touchdesigner-mcp by 8beeeaaat -- the first widely used TouchDesigner MCP server (2025). Envoy's earliest tool vocabulary (get_td_classes, get_td_class_details, get_td_info, get_module_help, exec_node_method, execute_python) follows the names that project established.
  • td-mcp-rs by asyade -- a Rust-daemon take on the same problem (2026). Its OS-dialog dismissal, per-call instance addressing, stable error codes with fix hints, any-family capture through an OP Viewer TOP, shader lint on DAT writes, and offline installation into a project file are ideas Embody adopted and built (see the roadmap).

If Embody carries an idea of yours and this list misses it, open an issue and it will be added.