MCP Tools Stack

Every other node in this pack's MCP family outputs one thing: a single mcp_tool. And the OpenAI Compatible LLM node takes exactly one mcp_tools list. Something has to bridge that gap, and that something is MCPToolsStack - the plumber of the family. It collects any number of MCP_TOOL inputs (remote, local, presets, all of them) into a single list you feed to the LLM node.

Without it you'd be stuck building workflows with exactly one tool. With it, you connect a Tavily search preset and a DeepWiki preset and a local time server to one stack, and the LLM sees all three. The README's example workflows all route through it, and honestly, most of the time you'll add it to a workflow because the chain demands it, not because you thought about it. That's what glue nodes are for.

How it works

The clever part is the dynamic inputs. You connect a tool to tool_1 and the node's frontend JS automatically adds a tool_2 input, then tool_3, and so on. The Python side accepts arbitrary tool_N slots so the UI can keep spawning them. There's also hidden backward-compat handling for old workflows that used a single tool or tools input - the stack sorts everything into a stable order and merges it into one list.

Two outputs: mcp_tools is the list you wire into the LLM node's mcp_tools input, and tools_json is a pretty-printed JSON dump of everything the stack collected. That second output is genuinely handy - when a tool isn't behaving, glance at tools_json to see what actually got built.

Installing it

Same pack, same story: ComfyUI Manager → search "OpenAI Compatible LLM", or

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

Restart. Light dependencies (requests, python-dotenv), nothing else to download. The dynamic-input JS ships inside the pack.

Where people get burned

Not much to trip on here, which is the point of a glue node. The one real gotcha: if you connect tools but the LLM never calls any of them, the problem is upstream - check that your api_base_url is /v1/responses (MCP tools require the Responses API) and that the LLM's prompt actually mentions the tools. The stack is faithfully doing its job; the signal is just lost before it gets to the model. Use tools_json to confirm the tools list is actually populated and look right before you go chasing anything else.