CLIOne CLI for the whole engine: generate images, video, audio, and 3D, run real workflow graphs, and pipe every output where you need it. Built for coding agents, CI jobs, and anyone with a terminal.
$
Bring ComfyUI workflows and creative models into the tools you already use. Run on Comfy Cloud or your own machine.
Python CLI · Cloud or Local · Agent skills included
Session transcript: comfy generate renders a cat and saves cat.png. comfy templates fetch retrieves image_z_image_turbo as zimage.json. comfy workflow set-slot changes the prompt to a lighthouse at dusk. comfy run submits that workflow to Cloud and comfy download saves f5308bc5_000.png. The animation ends with both saved images.
One command installs five bundled skills that teach Claude Code, Cursor, and any AGENTS.md-aware tool the whole CLI. Below are six production briefs. Every prompt is ready to paste.
Illustrative examples
Run our pack-shot workflow for every SKU in ./skus.csv, 4 angles each, route it to cloud, and download everything to ./shots.
Board a 30-second spot from BRIEF.md: 8 keyframes, consistent hero character, 16:9, contact sheet at the end.
Load our saved character workflow and generate a 4-view turnaround for each hero in ./characters.
Extend this plate to a 21:9 crop, matte-painting style. Keep the horizon line and the light direction.
Build a 3D hero prop from sketch.png, game-ready, and hand me the .glb.
Match frame_04.png's palette and produce campaign key art in 1:1, 4:5, and 16:9.
New models are added as they launch.
The contract lives in the CLI itself. No wrapper, no guesswork.
--json
Every command accepts --json and returns one envelope shape: ok, data, error, and a hint your agent can act on. Pipe it into anything that reads JSON.
comfy --json discover
One call dumps the entire command tree, output schemas, and error codes. Your agent reads the contract instead of guessing flags.
comfy run --workflow
Not one-shot endpoints. A run is a workflow JSON your agent can edit, validate, rerun, and hand to a teammate. It's the same graph that opens on the ComfyUI canvas.
The terminal counterpart to Comfy MCP: the same engine and the same account, scripted.
comfy-cli is open source, and so is the engine it drives. Install from PyPI, read the source when you need to, and never build on a black box.
Multi-step workflow graphs, thousands of community nodes, and partner models through the same commands. A creation engine, not a menu of hosted endpoints.
A run is a workflow file plus a seed. Check it into git, diff it in review, replay it in CI, and get the same output every time.
Every command auto-detects where to run: signed in, jobs route to Comfy Cloud GPUs; otherwise they hit the ComfyUI on your machine. Override per call with --where.
One account, two surfaces. Chat through Comfy MCP, script through the CLI, and let your agent pick the right one per task. Cloud outputs land in the same asset library either way.
01
pip install comfy-cli, then comfy setup. Sign in with OAuth for cloud GPUs, and comfy skills install teaches your agent the rest.
02
comfy --json discover hands your agent the full command tree; templates, models, and nodes are one search away.
03
Queue a run, wait on the job, download the outputs. Or pipe all three into a single line.
Over every model, every node, every step, every output.
Run ComfyUI on your own hardware.
SEE DESKTOP FEATURESThe full power of ComfyUI from anywhere.
SEE CLOUD FEATURESDeploy and scale ComfyUI workflows.
SEE PLATFORM FEATURESEnterprise-grade infrastructure for the creative engine inside your organization.
SEE ENTERPRISE FEATURESAny coding agent that can run shell commands. comfy skills install adds bundled skills for Claude Code, Cursor, and any AGENTS.md-aware tool, and the paste-prompt setup works in Codex, Gemini CLI, OpenClaw, Hermes, and the rest. For humans: any terminal, with tab completion via comfy --install-completion.
Both drive the same engine with the same account. Comfy MCP is for interactive agent work in chat: results come back in the session. The CLI is for automation: batch runs, CI, shell pipelines, and outputs saved to disk with --download or emitted as --json. Many teams run both and let the agent pick per task. Comfy MCP lives at https://comfy.org/mcp/.
The CLI is free and open source. With a Comfy account you can browse templates, models, and nodes at no cost, and runs on your own hardware are free. Partner models are the exception: they run on partner infrastructure and spend Comfy credits. Cloud runs spend credits too, either added directly or through an existing Comfy Cloud subscription.
Every command auto-detects where to run: signed in to Comfy Cloud, commands route to cloud; otherwise they run against your local server. Override per call with --where local or --where cloud, set the COMFY_WHERE environment variable, or persist a default with comfy set-default --where cloud.
Browser OAuth is the recommended path: comfy cloud login opens a sign-in, and the CLI refreshes tokens for you. API keys are the fallback for headless and CI use where no browser exists: create one at https://platform.comfy.org/profile/api-keys, then set COMFY_API_KEY or pass --api-key per call.
--download saves generate results to the path you name. comfy download pulls a run's outputs into ./outputs/, and it reads prompt IDs straight from piped stdin, so comfy run --workflow flux.json --wait | comfy download is the whole loop. On cloud, outputs also land in your Comfy Cloud asset library.
Yes. comfy-cli lives at https://github.com/Comfy-Org/comfy-cli and installs from PyPI with pip install comfy-cli, or from Homebrew. The ComfyUI engine it drives is open source too.
comfy generate is in beta: flag names, model aliases, and output formats may change while we iterate. The underlying partner endpoints are stable. File feedback on the comfy-cli GitHub repo.
Yes, that's what it's for. comfy setup --where cloud --api-key comfyui-... --non-interactive configures a runner without a browser, every command takes --json, and comfy validate --workflow catches unknown nodes and missing models before you spend compute.
• Generate images, video, audio, and 3D through partner models like Flux, Seedance, Kling, and Ideogram • Build, edit, validate, and run workflow graphs; save and rerun them • Queue batches, wait on jobs, and download every output • Search templates, models, and nodes with full schemas • Compose multi-stage workflows from typed fragments Run comfy skills install once, and the bundled skills teach your agent all of it.
