Comfy MCP and Sume MCP in one Claude Code session

Add both as remote HTTP MCP servers under different names, sign in to each, and tell the agent which one to use for what. Comfy's cloud endpoint is https://cloud.comfy.org/mcp and Sume's is https://mcp.sume.com/mcp. Claude Code's claude mcp add --transport http <name> <url> form is the one Sume's quickstart documents, and Comfy's docs say to add its URL as a remote MCP server in any client and sign in when prompted, read 2026-10-03.

Use two names so the tools stay separate: Claude Code names MCP tools mcp__<server>__<tool>, so comfy and sume give you two prefixes you can write permission rules against.

What are the commands?

Comfy's page lists OAuth for Claude Code and Codex, and an X-API-Key header for clients without OAuth. Sume uses OAuth for interactive clients, with a consent page on its MCP host.

claude mcp add --transport http comfy https://cloud.comfy.org/mcp
claude mcp add --transport http sume https://mcp.sume.com/mcp
claude mcp login sume
claude mcp list

How do the two servers compare?

Both servers are remote HTTP MCP endpoints with OAuth, and both put discovery first. The differences that matter are in billing and in what each one produces.

How do I keep the agent from mixing them up?

Say it in the prompt or a project instruction: Comfy for anything that must remain an editable workflow, Sume for finished stills, clips, audio and transcripts. Then check each connection by name with a read-only call: tools_list and mcp_health on Sume, and Comfy's free template search on Comfy. Both servers are discovery-first, so a wrong guess costs nothing until a paid tool is called.

Long clips need patience on the Sume side: one jobs_wait call holds for a bounded slice, and the agent should wait again rather than resubmit. The tool timeout post covers the settings.

A project instruction file is the right place for the routing rule, because it applies to every session without retyping. Keep it to two lines and name the servers exactly as you added them, comfy and sume. If the agent still picks the wrong one, run claude mcp list to confirm both servers are connected before blaming the prompt.

Does a second server cost context?

Yes: every connected server adds tool definitions to what the agent can see. Sume's own docs say to call tools_list for the session-visible subset, and a read-only OAuth session already sees fewer tools than a Write session. The tool search post and the tool limit post cover long tool lists in Claude Code and other clients.

What should I watch for?

Two servers mean two sets of rules. These are the places where a developer is most likely to get surprised.

• Two bills. Comfy Cloud generations need a Comfy subscription (or the free runs), and Sume jobs draw on a Sume wallet. A prompt that says "generate" does not say which.
• Write scope. Leave Write off on the Sume connection until a task needs a paid call; reads such as jobs_list and catalog_list work without it.
• Media hand-off. Sume's CLI docs call first-party result URLs public but still user data; treat them accordingly if you pass one into a Comfy workflow, and check Comfy's tools for how inputs are supplied.
• Partner models. Comfy's docs list partner models such as Flux, Grok, Gemini, OpenAI and Ideogram, and say they run on partner infrastructure and spend credits even locally.