mcp.film / Model Hubs / Comfy MCP
Comfy MCP
Comfy Org's own MCP: drive your local ComfyUI, or Comfy Cloud, from an agent.
What it does
The official server for ComfyUI, in two connections that share a vocabulary but not an identical tool list: a stdio server that drives the ComfyUI on your own machine through comfy-cli, and a hosted endpoint at cloud.comfy.org for people without a GPU. It reads your actual install — custom nodes, installed models, GPU specs — so an agent can tell you a model will not fit before it downloads 20GB, then run the workflow, poll the job and fetch the outputs. For a filmmaker that means local open video models on your own hardware with the vendor's own maintenance behind the server.
Connect
https://cloud.comfy.org/mcp
claude mcp add --transport http comfy-cloud https://cloud.comfy.org/mcp
{
"mcpServers": {
"comfy-mcp": {
"command": "npx",
"args": [
"-y",
"mcp-remote",
"https://cloud.comfy.org/mcp"
]
}
}
}
{
"mcpServers": {
"comfy-mcp": {
"url": "https://cloud.comfy.org/mcp"
}
}
}
Local install: pip install comfy-mcp "comfy-cli>=1.14.0" then: claude mcp add comfy-mcp -e COMFY_BIN=/path/to/venv/bin/comfy -- comfy-mcp
COMFY_API_KEY — https://platform.comfy.org/profile/api-keys
Sample tools
run_workflowgenerate_imagepartner_generaterun_templatejobfetch_outputsvalidate_workflowsearch_templatessearch_modelsserver_infolaunch_comfyuiinstall_nodeField notes
Comfy Org open-sourced the local server on Aug 18 2026. Re-checked 11 Oct 2026: the repo is busy — last commit Oct 10, 269 stars, not archived — but PyPI `comfy-mcp` and the newest tag are both still 0.10.0, so `pip install` gets you August's build and the fixes since then only arrive from a source checkout. Main now requires the MCP Python SDK 2.x (`mcp>=2,<3`), which matters if you install it into an environment that pins 1.x. Read the licence before you ship it inside a product: AGPL-3.0-or-later OR Commercial, not the MIT most of this directory carries. The local route is not a one-liner — it needs Python 3.10+, comfy-cli 1.14.0+ on PATH and a ComfyUI workspace, and the server refuses to start against an older comfy-cli. Local runs are free on your own hardware; the partner models (Flux, Ideogram, Kling, Veo, Runway and ~30 others) bill credits and each spend needs explicit confirmation. Cloud is the other half: `https://cloud.comfy.org/mcp` authenticates over OAuth in Claude Code and by an X-API-Key header (keys start `comfyui-`) for Cursor and headless clients; discovery works on a free Comfy account, but running generations needs an active Comfy Cloud subscription. The cloud tool list is its own — `submit_workflow`, `get_job_status`, batches, saved workflows, shareable app links — rather than a mirror of the local one. Caveats: it is a stdio subprocess, so it cannot coordinate an external LLM runtime, and pointing it at a remote ComfyUI disables the lifecycle tools (launch/stop/update). The older community `comfyui` entry still works but has not moved since January.
Also in Model Hubs
Drive a local ComfyUI: 40+ tools for image, video, audio and 3D workflows.
Hedra's own hosted MCP: one OAuth URL onto its whole image, video and audio catalogue.
WaveSpeed's live image, video, audio and 3D catalogue behind seven tools, hosted or via npx.
One hosted endpoint over 450 models: text-to-video from Veo, Kling, Seedance and Hailuo on one pay-as-you-go balance.
Official hosted hub: run any image, video, audio or 3D app, then poll the task.
Browse, price, and run 130+ image, video, audio and voice models through one MCP.
Pairs well with
Hosted MCP for AKOOL face swap, character swap, image-to-video, TTS and avatar lookup.
Studio MCP for talking avatars, locked recurring characters, and batched output.
Official bitHuman CLI server: build avatar agents from a face image and render WAV-driven talking video locally.