touchdesigner-mcp
MCP server that lets Claude drive a running TouchDesigner instance to create operators, wire them, set parameters, run arbitrary Python, and introspect the td API.
README
touchdesigner-mcp
Model Context Protocol (MCP) server that lets Claude drive a running TouchDesigner instance — create operators, wire them together, set parameters, run arbitrary Python, and introspect the td API.
Architecture
Claude (MCP client) ──stdio──► touchdesigner-mcp (Python) ──HTTP POST──► Web Server DAT ──► td.run() ──► main-thread eval/exec
The MCP server is a thin stdio bridge. All TD mutation happens on TouchDesigner's main thread via a Web Server DAT callback that execs or evals the code you send.
Prerequisites
- TouchDesigner (any recent 2022+ build — tools introspect the live API)
- Python 3.10+
uvor plainpip
Install via .mcpb (recommended)
The .mcpb bundle is a single-file install for Claude Desktop. It vendors the Python deps and ships the TD-side .tox alongside, so both halves come from one file.
- Build the bundle from the repo root:
bash scripts/build_mcpb.sh # → built dist/touchdesigner-mcp.mcpb - Double-click dist/touchdesigner-mcp.mcpb. Claude Desktop opens a config form — defaults are fine for local TD (
TD_HOST=127.0.0.1,TD_PORT=9980,TD_PATH=/mcp). Leave the multi-instance fields blank unless you're targeting more than one TD process. Click Install. - Continue with the TouchDesigner-side setup below, then skip the "MCP client configuration" section (the bundle wires that for you).
Requires python3 on PATH (default on most macOS/Linux installs; install from python.org if missing).
Install from source
git clone https://github.com/mrinalghosh/touchdesigner-mcp
cd touchdesigner-mcp
uv venv
source .venv/bin/activate
uv pip install -e .
This exposes a touchdesigner-mcp console script inside .venv/bin/.
TouchDesigner-side setup (one-time, per .toe)
- Open your project in TouchDesigner.
- Drag td_component/touchdesigner_mcp.tox into
/project1(or any persistent COMP). The component ships a pre-wired Web Server DAT on port9980with the callbacks already attached. - Save the .toe.
<details> <summary>Manual setup (if you'd rather not use the .tox)</summary>
- Inside
/project1create a Web Server DAT (suggested name:mcp_webserver). - Set Port =
9980, Active = On. - Replace the auto-created Callbacks DAT contents with td_component/webserver_callbacks.py.
- Save the .toe.
</details>
Verify from a terminal:
curl -s -X POST http://127.0.0.1:9980/mcp \
-H 'content-type: application/json' \
-d '{"code": "_result = app.version", "mode": "exec"}'
# → {"ok": true, "result": "2023.xxxxx"}
Or run the packaged smoke test with TD open:
uv run python scripts/smoke_test.py
MCP client configuration
Claude Desktop
Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) — single-instance form:
{
"mcpServers": {
"touchdesigner": {
"command": "/absolute/path/to/touchdesigner-mcp/.venv/bin/touchdesigner-mcp",
"env": {
"TD_HOST": "127.0.0.1",
"TD_PORT": "9980"
}
}
}
}
Multi-instance form (target several TD processes from one Claude session):
{
"mcpServers": {
"touchdesigner": {
"command": "/absolute/path/to/touchdesigner-mcp/.venv/bin/touchdesigner-mcp",
"env": {
"TD_INSTANCES": "main=127.0.0.1:9980,fx=127.0.0.1:9981,stage=192.168.1.40:9980",
"TD_DEFAULT_INSTANCE": "main"
}
}
}
}
Restart Claude Desktop after editing.
Claude Code CLI
claude mcp add touchdesigner /absolute/path/to/touchdesigner-mcp/.venv/bin/touchdesigner-mcp \
--env TD_INSTANCES=main=127.0.0.1:9980,fx=127.0.0.1:9981 \
--env TD_DEFAULT_INSTANCE=main
See claude_desktop_config.example.json for both forms in one place.
Environment variables
| Var | Default | Purpose |
|---|---|---|
TD_HOST |
127.0.0.1 |
Single-instance host |
TD_PORT |
9980 |
Single-instance port |
TD_PATH |
/mcp |
HTTP path the Web Server DAT answers on |
TD_INSTANCES |
— | Multi-instance map: name=host:port[/path],... (overrides the single-instance vars) |
TD_DEFAULT_INSTANCE |
first in map | Which instance unqualified tool calls target |
TD_TIMEOUT |
10.0 |
HTTP timeout in seconds |
Per-instance path override: TD_INSTANCES="main=127.0.0.1:9980/mcp,dev=127.0.0.1:9981/mcp-dev".
Tools
Every tool accepts an optional instance argument to target a specific TD process. Omit it to hit TD_DEFAULT_INSTANCE.
Meta
list_instances— show configured TD processesping/ping_all— health check
Arbitrary code
exec_python(code)— runs on TD's main thread; assign to_resultto return a valueeval_python(expression)— single-expression eval
Node lifecycle
create_operator(parent_path, op_type, name)delete_operator(path)rename_operator(path, new_name)move_operator(path, x, y)
Parameters
set_parameter(path, param, value)get_parameter(path, param)list_parameters(path)pulse_parameter(path, param)bind_parameter_expression(path, param, expression)— switch a parameter to Expression mode with verification (reports any eval exception or new op error)
Wiring
connect_operators(source_path, target_path, source_output=0, target_input=0)disconnect_input(path, input_index=0)
Query
list_children(comp_path)find_operators(root_path='/', op_type=None, name_pattern=None, depth=4)get_errors(path='/', recurse=True)
Templates (multi-op recipes that encode known gotchas)
list_templates— discover available templates and their optionscreate_from_template(template, parent_path, name_prefix, options=None)— instantiate one. Ships withchop_source_with_null,glsl_top_vec4_uniform,audio_in_with_analyze,feedback_loop_top,render_pipeline.
Introspection (lets the model discover the live td API rather than guessing)
get_td_info— version, project, Python, platformget_td_classes(name_contains=None)get_td_class_details(class_name)get_module_help(name)
Example prompts
"Under
/project1, create a Noise TOP calledn1and a Level TOP calledlvl1, wire n1 → lvl1, then setlvl1.Brightnessto 0.6."
"List every TOP under
/project1and report which ones have errors."
"Show me
td.noiseTOP's parameters so I know what I can tweak."
Security
The Web Server DAT callback executes arbitrary Python against your live project. Bind it to 127.0.0.1 only, never expose the port to the internet, and don't run untrusted prompts against a TD instance with valuable state open.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器