Local Models & Open Clients

Envoy is a standard MCP server — any MCP-compatible client can drive TouchDesigner through it, including open clients running local models. This page covers the two clients that matter most for that workflow — OpenCode and LM Studio — plus practical model recommendations and the settings that make small models reliable against Envoy's tool surface.

Everything runs on your machine: TouchDesigner, the model, and the agent. No account, no API key, no cloud.

OpenCode

OpenCode is the most widely adopted open coding agent, with first-class support for local models through LM Studio, Ollama, llama.cpp, or any OpenAI-compatible endpoint.

Automatic setup (recommended)

Set the AI Client parameter on the Embody COMP to opencode (or pick OpenCode in the setup wizard). Embody generates opencode.json in your project root with:

• mcp.envoy — spawns the same STDIO bridge Claude Code uses, so OpenCode gets the bridge meta-tools (get_td_status, launch_td, restart_td, switch_instance), a cached tool list while TouchDesigner is closed, automatic reconnection, and instance-registry identity checks.
• instructions — loads the generated .claude/rules/*.md alongside AGENTS.md. (OpenCode reads AGENTS.md natively and discovers .claude/skills/ through its Claude-compatibility layer, so the full rule-and-skill set carries over with one copy on disk.)
• A permission block matching your Tool Permissions posture (only when the file is created fresh — an existing opencode.json is merged into, never overwritten).

opencode.json embeds absolute paths, so Embody gitignores it and mirrors it into sibling *-wt-* worktrees, exactly like .mcp.json.

Manual setup

Point OpenCode at a running session directly (simplest, but no bridge features — the connection dies when TouchDesigner closes):

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "envoy": {
      "type": "remote",
      "url": "http://127.0.0.1:9870/mcp",
      "enabled": true
    }
  }
}

Use 127.0.0.1, not localhost — on Windows, localhost can resolve to IPv6 first and stall every request for ~2 seconds.

Connecting a local model

OpenCode configures local providers in the same opencode.json. LM Studio example (from the OpenCode docs):

{
  "provider": {
    "lmstudio": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LM Studio",
      "options": { "baseURL": "http://127.0.0.1:1234/v1" },
      "models": {
        "qwen/qwen3-coder-30b": { "name": "Qwen3 Coder 30B" }
      }
    }
  }
}

Ollama is the same pattern with baseURL: http://127.0.0.1:11434/v1; llama.cpp's llama-server with :8080/v1.

LM Studio as an MCP host

If you'd rather not use a terminal at all: LM Studio (0.3.17+) can host MCP servers directly in its chat UI — a local model with Envoy's tools, no CLI. Add this to LM Studio's mcp.json (Program tab → Install → Edit mcp.json):

{
  "mcpServers": {
    "envoy": {
      "url": "http://127.0.0.1:9870/mcp"
    }
  }
}

LM Studio asks for confirmation before each tool call by default. Note its mcp.json is global to LM Studio (not per-project), so this connects whichever TouchDesigner project's Envoy is on that port.

Which model? (snapshot: July 2026)

Tool-calling quality matters more than raw parameter count, and MoE models with small active-parameter counts are the key to good speed on modest hardware — a dense 27–31B model that spills into system RAM crawls, while an 80B MoE with 3B active parameters stays fast. Current picks:

Hardware	Models that work well
16 GB VRAM laptop	gpt-oss-20b (fits fully, ~60K context), GLM-4.7-Flash (30B-A3B), Gemma 4 26B-A4B
16 GB VRAM + 64 GB RAM	gpt-oss-120b or Qwen3-Coder-Next (80B-A3B) via llama.cpp --n-cpu-moe — big-model quality at ~15–25 tok/s
24–32 GB VRAM	Qwen3.6-35B-A3B or Qwen3-Coder-30B-A3B fully resident; Devstral Small 2 (24B dense)
96 GB (RTX PRO 6000-class)	gpt-oss-120b fully in VRAM (~130+ tok/s), Qwen3-Coder-Next at high quants
Apple Silicon 64–128 GB	GLM-4.5-Air, gpt-oss-120b (MLX)

Qwen3-Coder-Next is the current local benchmark leader for agentic coding (70.6% SWE-bench Verified, Apache 2.0). For creative TouchDesigner direction — as opposed to mechanical code edits — bigger models are noticeably better; 12B-class models handle Python well but plateau on aesthetic judgment.

Cheap cloud fallbacks that pair well with a local setup: GLM-4.7-Flash is free on Z.ai's API, DeepSeek V4 is $0.30/M input tokens, and OpenRouter's :free models allow 50 requests/day (1,000/day once you've ever bought $10 of credit).

Settings that make small models reliable

These four fix the vast majority of "my local model won't use the tools" reports:

1. Context window ≥ 16K, ideally 32K. Ollama's default context is 4,096 tokens and it truncates silently — the system prompt plus tool schemas already exceed that, so the model literally never sees Envoy's tools. Set num_ctx/OLLAMA_CONTEXT_LENGTH (or the equivalent in LM Studio) to 16–32K minimum.
2. Quantization Q5/Q6 or better. Q4 quants degrade tool-call JSON validity and multi-step consistency faster than chat quality suggests. KV cache at q8 is essentially free; avoid q4 V-cache for agent work.
3. llama.cpp needs --jinja. Without it, tool-call delimiters pass through as literal text.
4. Trim the tool surface. Envoy exposes 50+ tools ≈ 15–25K tokens of schemas — on a 32K local context that's most of the window gone, and research shows tool selection accuracy degrades sharply as tool count grows. In OpenCode, the top-level tools map controls which tool schemas reach the model (the permission map only governs call approval — a denied tool's schema still occupies context). Disable the full surface with a glob, then re-enable a lean core:

{
  "tools": {
    "envoy_*": false,
    "envoy_get_td_status": true,
    "envoy_query_network": true,
    "envoy_read_tdn": true,
    "envoy_get_op": true,
    "envoy_get_op_errors": true,
    "envoy_get_parameter": true,
    "envoy_set_parameter": true,
    "envoy_create_op": true,
    "envoy_connect_ops": true,
    "envoy_delete_op": true,
    "envoy_batch_operations": true,
    "envoy_execute_python": true,
    "envoy_get_network_layout": true,
    "envoy_set_op_position": true,
    "envoy_capture_top": true,
    "envoy_get_dat_content": true,
    "envoy_edit_dat_content": true,
    "envoy_get_docs": true
  }
}

That 18-tool core covers inspect, build, wire, position, verify, and capture — enough for real work while leaving the context to the model.

Expectations

A well-configured local setup handles operator work, parameter edits, Python, and network building competently. Long multi-step autonomy and creative direction still favor larger models — many people run local for iteration and switch to a hosted model for the hard steps. Both connect to the same Envoy; nothing about the project changes.