Comfy Cloud
Comfy Cloud runs ComfyUI in the cloud and exposes it to agents as an MCP server: image/video/audio/3D generation via partner models (Ideogram, Flux, Kling, Veo, Seedance, and more) and open-source pipelines, plus full ComfyUI workflow authoring — search templates, override inputs, submit custom graphs, and fetch outputs as signed URLs.
Scope: this skill connects to the hosted Comfy Cloud MCP server and requires a Comfy account — it does not drive a local ComfyUI install. A local-ComfyUI MCP is on the roadmap and will ship as a separate skill; until then, for local instances see comfy-cli.
Connect
The server supports two auth methods: OAuth (browser sign-in, nothing to manage — recommended) and API keys (for headless or CI setups).
Option A — OAuth (recommended)
bashopenclaw mcp set comfy '{"url":"https://cloud.comfy.org/mcp","transport":"streamable-http","auth":"oauth"}' openclaw mcp login comfy openclaw gateway restart
openclaw mcp login prints an authorization URL — open it, sign in to your
Comfy account, and finish the flow as the CLI directs (it may ask you to rerun
with --code <code>). Verify with openclaw mcp status --verbose. Note that
OpenClaw ignores static headers on a server entry once auth: "oauth" is
set, so don't combine the two.
Option B — API key (headless / CI)
-
Get an API key — sign in at cloud.comfy.org, then create a key under Settings → API Keys. Keys start with
comfyui-. -
Export it (put this in your shell profile or OpenClaw's env):
bashexport COMFY_API_KEY="comfyui-..." -
Register the MCP server with OpenClaw:
bashopenclaw mcp set comfy '{"url":"https://cloud.comfy.org/mcp","transport":"streamable-http","headers":{"Authorization":"Bearer ${COMFY_API_KEY}"}}' openclaw gateway restartThe server also accepts the key as an
X-API-Key: ${COMFY_API_KEY}header, but prefer theAuthorization: Bearerform above — standard headers survive every client's proxy layer, and some OpenClaw builds have dropped custom headers on streamable-http transports (see openclaw#65590). If you get a401with a key you know is valid, you are probably hitting that: switch to the Bearer form, route through mcporter, or just use OAuth (Option A).
Either way: browsing/search tools work with any account; generation consumes Comfy Cloud credits and needs a subscription or credit balance.
Which tools to use
Images and video from a named model (Ideogram, Flux, Gemini, Kling, Veo,
Seedance, …): before calling partner_generate, check whether the named
family also has an OSS route — some families (MiniMax H3 is a current
example, as of 2026-08) ship BOTH a paid partner node and open-source weights
under the identical display title. Run search_templates for the family name
alongside search_nodes; if both exist, tell the user the OSS option exists
(it has no partner/API fee, but running it still spends ordinary Comfy Cloud
compute credits — it isn't free on Comfy Cloud, only free of the partner
surcharge) and ask which they want. If only the partner route exists, or the
user picks paid, call partner_generate — it runs the provider through Comfy
Cloud and saves the result to your asset library. Do not hand-build a workflow
for plain text-to-image/video when a partner model is named — unless the user
asked to build or create a workflow: partner_generate returns only the
output, never a workflow, so build the graph around the partner node instead.
Anything workflow-shaped (open-source models, LoRA/ControlNet, multi-step
pipelines): start from a template — search_templates → get_template_schema
→ run_template with input_overrides. For fully custom graphs use
submit_workflow (the graph must end in an output node like SaveImage, or
validation rejects it). Give every node in a custom graph a _meta.title
naming its role in plain words ("Load driving video", "Remove background",
"Save video") — the canvas shows these titles; without them the user sees only
class names. Save and reuse with save_workflow / run_saved_workflow. When
the user asked to build or create a workflow (not just for an output), save it
with save_workflow after the run, then hand them the link from
get_workflow_canvas_url and mention the saved layout is auto-generated and
may need tidying.
Editing an existing image: upload_file first, then reference the
uploaded file from a LoadImage node (or pass it to partner_generate for
image-to-image partners). Never inline base64 image data into a workflow.
Getting results: wait_for_job → get_output. Outputs come back as
time-limited signed URLs — download them exactly as returned (don't edit query
params, they're part of the signature). get_queue and cancel_job manage
in-flight jobs; submit_batch / wait_for_batch fan out variations.
Discovery: search_models (checkpoints/LoRAs on the platform),
search_nodes (available ComfyUI nodes with wiring hints),
get_prompting_guide (model-specific prompting advice).
Example prompts
- "Generate an image of a cozy bookstore café on a rainy afternoon with Ideogram and download it to ./bookstore.png"
- "Take ./product.jpg, remove the background, and upscale the result 2×"
- "Find a text-to-video template for Kling, run it with the prompt 'a paper boat drifting down a rain gutter, cinematic', and give me the output URL"
Notes
- Generation spends Comfy Cloud credits; search/discovery tools are free.
- Full MCP docs and per-client setup: docs.comfy.org/cloud/mcp
