Claude MCP Manager

Claude MCP Manager is the pack's control panel for Model Context Protocol servers - the plumbing that lets Claude Code reach outside its sandbox and use external tools. MCP is how the pack's own Reddit scraper gets a browser: the scraper talks to a Playwright MCP server. This node is the general-purpose version of that wiring, letting you list what MCP servers you have, add or remove them, and pull example configs, all from inside the graph instead of a terminal.

The inputs

The action dropdown is the whole UI, really:

• list - shows your configured MCP servers, via claude mcp list --json. list_format switches between just names, a full breakdown (command and args), or raw json.
• enable - registers a server using mcp_name plus the mcp_config JSON you provide, via claude mcp add-json --scope user. This is where you'd register your own Playwright-style server with its command, args, and env.
• disable - removes a server by name (claude mcp remove).
• config - doesn't change anything; it just prints example configurations for Slack, browser-tools, Notion, and Figma to give you a template for the JSON format.

Outputs

Outputs: mcp_info (the human-readable result text), mcp_config (an MCP_CONFIG value - internally a marker like enabled:NAME or disabled:NAME that Execute can read), and mcp_data (the raw JSON list of servers for the list action).

Installing

Install: ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/christian-byrne/claude-code-comfyui-nodes

and restart. It needs the claude CLI signed in, same as the rest of the pack, since it's all CLI calls under the hood.

Common issues

Two honest caveats. First, the mcp_config you feed enable has to be a valid server definition - a JSON object with at least a command (like "command": "playwright-mcp-server"), or the add fails. And if the underlying server binary isn't installed, the registration "succeeds" but the server never actually launches when Claude tries to use it; the node can't install third-party tools for you. Second, the Execute node's handling of whatever you wire in here is deliberately thin - it mostly checks whether the marker says enabled or disabled and logs errors, rather than doing deep MCP orchestration. The heavy MCP lifting in this pack lives inside the Reddit scraper, which auto-configures Playwright itself. Treat this node as the manual override and inventory tool: great for seeing what's configured and adding servers by hand, less of a magic bullet than the name suggests. If you just want the browser-scraping workflow, skip this node entirely and let the scraper handle it.