comfy
Generate images, video, audio, and 3D with Comfy Cloud — search hundreds of models and workflow templates, run custom ComfyUI workflows, and manage generation jobs through the hosted Comfy Cloud MCP server. Cloud-only — connects to the hosted service, not a local ComfyUI install. Use for any "generate/edit an image", "make a video", "text to 3D", or ComfyUI workflow task.
How do I install this agent skill?
npx skills add https://github.com/comfy-org/comfy-skills --skill comfyIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill allows the agent to interact with the Comfy Cloud service for media generation. It uses standard authentication methods (OAuth or API keys) and connects exclusively to official ComfyUI domains. No security risks were identified.
- Socketpass
No alerts
- Snykwarn
Risk: MEDIUM · 1 issue
What does this agent skill do?
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)
openclaw 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):
export COMFY_API_KEY="comfyui-..." -
Register the MCP server with OpenClaw:
openclaw 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 (overwrite: false; new name on collision) after the
run, then hand them the link from get_workflow_canvas_url using the
workflow_id from the save result (from list_saved_workflows if the result
omits it — never the filename, which publishes a public snapshot); on a
run_template route, get_template, apply the run's input_overrides to
that body and save the result, so the canvas matches the run; after any
from_prompt_id reruns, layer each rerun's input_overrides on top in run
order before saving, so the canvas matches the latest run. 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
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/comfy-org/comfy-skills/comfy">View comfy on skillZs</a>