MCP DeepWiki Remote Tool

So you've got an LLM node inside ComfyUI writing your prompts - good. Now it's mid-prompt and it needs to know something. Not "what rhymes with firefly," but "what does ComfyUI's own node API actually accept." That's the gap this node fills: it hands your LLM a live connection to DeepWiki's public MCP server, which is the AI-generated wiki for basically every public GitHub repository that matters.

DeepWiki is the thing Cognition runs - you point it at a repo, it reads the whole codebase and builds a browsable, Q&A-able documentation site. No API key, free, public. This node is a preset that wraps DeepWiki's Remote MCP server so the LLM in your workflow can ask questions about any public repo and get real answers, not hallucinated ones.

How it works

This node doesn't call DeepWiki itself. It builds one Remote MCP tool definition - a {"type": "mcp", "server_label": "deepwiki", "server_url": "https://mcp.deepwiki.com/mcp", "allowed_tools": [...], "require_approval": "never"} object - and hands it to the OpenAI Compatible LLM node. The API server on the other end (LM Studio, Ollama, OpenAI, whoever) is the one that actually opens the connection to DeepWiki when the model wants to ask something.

The catch, and it's a real one: Remote MCP tools only work over the Responses API. Your api_base_url needs to end in /v1/responses, not the plain /v1. Point it at the chat completions endpoint and the tools silently do nothing.

The inputs that matter

Only two, and you can usually leave them both alone:

• allowed_tools - defaults to ask_question and read_wiki_structure, the two read-only tools from OpenAI's own DeepWiki examples. Both are safe to expose; they only read. Leave it narrow unless you have a reason not to.
• server_url - defaults to https://mcp.deepwiki.com/mcp. You'd only touch this if the endpoint ever changes.

The single output, mcp_tool, is what you wire forward. The chain is always: MCP DeepWiki Remote Tool → MCP Tools Stack → OpenAI Compatible LLM.

Installing it

This ships in the godmt/comfyui-openai-llm pack. Easiest path is ComfyUI Manager - search for "OpenAI Compatible LLM" and install. Or the manual route:

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

Restart ComfyUI. That's it - the pack's real dependencies are just requests and python-dotenv. There's nothing heavy here and nothing to download; DeepWiki needs no key from you.

Where people get burned

The classic failure is building the chain, running the workflow, and getting a plain text answer with zero repo research. Checklist: /v1/responses in the URL, the DeepWiki node actually connected through a stack, and ask_question still in the allowlist. If the model just never calls the tool, nudge the prompt - models don't spontaneously use tools you didn't mention. One honest aside: DeepWiki's wiki for your repo may take a few minutes to build on their end the first time, and the answer quality tracks how good that wiki is. You're not doing anything wrong when it's mediocre on an obscure repo.

A word you should internalize for every MCP node in this pack: Remote MCP tools are sent with require_approval: "never" because ComfyUI has no approval UI, and MCP servers can read files or hit external services. DeepWiki is read-only and public, so this is the least scary node in the family. The LLMVISION disaster taught the ecosystem how much damage a malicious node can do - trust the pack, but only connect MCP servers you actually recognize.