MCP Local Command Tool

This is the node in the pack that gives your LLM real hands - and it's the one you should treat with the most respect. The Remote MCP nodes hand the model a connection to a server somewhere out there. This one launches a local MCP server on your machine via npx, uvx, or python, lets the LLM call its tools, then tears it down. Think: the model checks the actual time, reads a file, runs a calculation - the boring, factual stuff that LLMs are hopeless at from memory alone.

Concretely, the pack's own example is a time server: run uvx mcp-server-time --local-timezone=Asia/Tokyo and let the LLM ask it for the current time in Tokyo instead of guessing. Same idea scales to anything with an MCP stdio server - file tools, browser tools, whatever you trust.

How it works

During an execution the OpenAI Compatible LLM node takes the local MCP definition and runs a real handshake: it spawns the command, does the MCP initialize dance over stdin/stdout, lists the available tools, converts them to function tools the LLM can call, executes tools/call when the model asks, and feeds results back. Then it shuts the process down - the pack explicitly tries to leave no child processes behind, with process-tree termination as a Windows fallback.

The inputs that matter

• command - the launcher: npx, uvx, python, or an absolute path. This is a big one because it's arbitrary code execution by definition. The LLM is going to run whatever this launches.
• args_json - the arguments as a JSON array. Example: ["mcp-server-time", "--local-timezone=Asia/Tokyo"].
• server_label - a short stable prefix for the exposed tool names, e.g. time__get_current_time.
• allowed_tools - leave empty to expose everything the server has, or list the few you want. I'd strongly recommend listing.

Worth setting: startup_timeout_sec (default 15, for slow first-time uvx/npx downloads) and tool_timeout_sec (default 60, per tool call). env_json passes extra environment variables to the spawned process.

Output is a single mcp_tool, wired MCP Local Command Tool → MCP Tools Stack → OpenAI Compatible LLM, with the LLM node on /v1/responses.

Installing it

Same pack as the rest: ComfyUI Manager, search "OpenAI Compatible LLM", or

cd ComfyUI/custom_nodes
git clone https://github.com/godmt/comfyui-openai-llm.git

then restart. Dependencies are light (requests, python-dotenv) - but the command you run is a dependency too. uvx needs uv installed (pip install uv), and the first run of any npx/uvx server downloads its packages, which is why the timeout matters.

Where people get burned

Test the command in a terminal first. The README says it plainly: if uvx mcp-server-time --local-timezone=Asia/Tokyo fails in your shell, it fails in ComfyUI too, and you'll be debugging a dark node instead of a clear error. Other classics: forgetting /v1/responses so tools never appear, and an LLM that won't call a tool because the prompt never mentions it.

Then the security bit, said once and loudly: this node runs arbitrary commands on your machine with your user privileges. The ecosystem's own history - the LLMVISION infostealer that shipped as a "helpful" node and ended in a federal conviction - is the cautionary tale. Use it with servers you trust and allowlists you've narrowed. This is the highest-risk node in the pack, and it's not close.