Local copilot (no API)

tdmcp ships a local copilot so you can drive TouchDesigner with a free LLM running on your own machine — no paid API, no account, works offline. The command tdmcp chat opens a small chat page in your browser, wired to the same TouchDesigner bridge the other clients use.

It's the budget-friendly, private path: great for the everyday stuff, and it hands off to Claude or Codex the moment you want a whole system built.

Which path is this?

Claude Desktop is the no-terminal route. This page is for running tdmcp with a local model instead of a paid assistant — it needs Node.js 20+, like the Codex and Cursor paths.

What it's good for

The local copilot is given a curated, safe subset of the tools, so it's quick and hard to misuse. It's meant for the easy stuff:

• Inspecting your project — what's there, how it's wired.
• Reading errors and explaining what's wrong.
• Creating, wiring and tweaking individual operators — one node at a time.

It deliberately can't build whole systems (no Layer-1 generators) and can't run raw Python. When you want a full audio-reactive or generative network, click Escalate ⇪ in the UI: it copies a ready-to-paste prompt you hand to Claude or Codex. They drive the same project, so nothing has to move.

What you need

• TouchDesigner with the bridge on (the same one-line step as every client — see below).
• Node.js 20+ — used to launch the copilot.
• Ollama — the free local model runner. tdmcp chat starts it for you if it isn't already running.

Start it

The quickest path needs no clone — just Node and Ollama installed:

# one-time: install Ollama from https://ollama.com, then optionally pre-pull a model
ollama pull qwen2.5:3b      # optional — the UI also has a one-click pull

npx -y @dpantani/tdmcp chat # opens http://127.0.0.1:4141 in your browser

If you already cloned and built tdmcp (the from-source path), the command is simply tdmcp chat (or node dist/index.js chat).

tdmcp chat starts Ollama for you if the daemon isn't up — detached and left running, so closing the chat never takes your model offline. Useful flags:

• --no-ollama — don't auto-start it (for a remote endpoint or a daemon you manage yourself).
• --no-open — don't open the browser automatically.
• --help — list everything.

Which local model?

qwen2.5:3b is the default — benchmarked at 100% tool-calling on the simple-task workload, as reliable as bigger models but faster and lighter. Sub-3B models are flaky; bump to qwen2.5:7b only if you want more answer-quality headroom. More detail in the CLI reference.

Using the chat

The browser UI is wired to your live TouchDesigner project. It has:

• A read-only toggle — let it look but not touch.
• Live model switching and endpoint settings, plus a one-click model pull if a model isn't downloaded yet.
• Persistent history, so your conversation survives a restart.
• Escalate ⇪ — copies a handoff prompt for Claude or Codex when a task is too big for the local model.

Point it at a different model

By default the copilot talks to local Ollama, but it speaks the standard OpenAI-compatible API — so you can aim it anywhere with two environment variables:

Variable	Default	Use it for
TDMCP_LLM_BASE_URL	http://127.0.0.1:11434/v1	LM Studio, a cloud GPU, or a paid API.
TDMCP_LLM_MODEL	qwen2.5:3b	Any model id available at that endpoint.

Full list (including TDMCP_LLM_API_KEY and the chat port) is in environment variables.

Turn on the bridge

Like every client, the copilot needs the small bridge running inside TouchDesigner. Open the Textport (Dialogs → Textport and DATs), paste this one line, and press Enter:

import urllib.request; exec(urllib.request.urlopen("https://raw.githubusercontent.com/Pantani/tdmcp/main/td/bootstrap.py").read().decode())

You should see [tdmcp] bridge running on port 9980. See Install for details and how to remove it later.

Not connecting?

• Confirm the bridge is on: curl http://127.0.0.1:9980/api/info should return JSON.
• Make sure Ollama is installed and a model is pulled (the UI's model pull does this for you).
• Full Troubleshooting covers the common cases.

With TouchDesigner open and the bridge on, ask in plain language — "what's in this project?", "why is this node red?", "add a blur after the noise." For bigger ideas, see the prompt cookbook or escalate to Claude / Codex.