Comfy API Turns a ComfyUI Workflow Into a Production Endpoint

There is a specific kind of frustration familiar to anyone who has built something useful in ComfyUI. The workflow works on your machine, tuned over weeks, with the right custom nodes, the right LoRA, the right pinned version of everything. Then someone asks you to make it available to the rest of the team, or to a product, and the work you thought you had finished turns out to be half done.

Getting a ComfyUI workflow into production has traditionally meant rebuilding its environment somewhere else. You rent GPUs, install every custom node and model again, untangle Python dependencies until they stop fighting, and write the scaling logic around it. The graph is the same, but the engineering is entirely new, and it is the kind of work that has nothing to do with why you built the workflow in the first place.

ComfyUI launched Comfy API at the end of September to close that gap. It is live for everyone on a paid Comfy plan, and it does what the name suggests: it turns a ComfyUI workflow into an autoscaling API endpoint without you changing the workflow.

Package once, deploy on the GPU you choose

The mechanism is worth understanding because it is where most of the value sits. You start with a workflow JSON, or export a snapshot from Comfy Desktop. A builder reads the workflow, finds the models and custom nodes it needs, and helps resolve conflicting Python dependencies. You can override any version it picks.

That resolution step gets captured into a Build: the ComfyUI version, custom nodes, models, LoRAs and Python dependencies your workflow expects, all pinned together. From a Build you cut an immutable release and deploy that release as a managed endpoint with its own URL.

The immutable release is the important part. Because the release does not change after it is cut, the environment you tested is the environment you deploy. When you need to make a change, you update the Build and cut a new release, without touching the one that is live. Anyone who has watched a working pipeline break because a dependency drifted underneath it will recognise why this is more than a convenience.

Deployment runs on ComfyUI's Developer Platform, where builds, deployments, usage, spend and API keys live in one console. You can work in the browser, from the terminal, or with a coding agent. The command-line path is compact. Run `comfy build init` and it scans the ComfyUI install for custom nodes, models and pinned dependencies; run `comfy build push --release` to package and publish. There is also a copy-paste agent prompt on the Comfy API page for anyone who would rather hand the whole thing to a coding assistant.

The pricing, and who it is actually for

Workloads scale to zero when idle, which means no GPU time is billed when no workers are running. For bursty creative pipelines, that is the difference between a fixed monthly GPU bill and paying only for the work performed. Teams with latency-sensitive requests can keep workers warm instead, trading cost for the absence of cold starts.

GPU billing is per second, with published rates: an RTX PRO 6000 at $4.54 an hour, an H100 at $6.23, an H200 at $7.71, and a B200 at $11.23. A paid plan, Standard and up, is required, and usage is billed separately.

The most honest thing about the launch is what ComfyUI says about who should not use it. The team notes that if RunPod or Modal already works for you, keep using it, and frames Comfy API as a fit for when managing the ComfyUI environment itself is the headache. That is a more useful positioning than the usual everything-platform pitch, and it tells you the target customer: someone whose bottleneck is dependency management, not raw compute procurement.

Why this changes what a small team can ship

The clearest way to see the shift is through the handoff problem. The person who builds a workflow is often not the only person who needs to run it. A designer tunes a product-photography workflow with custom nodes and a fine-tuned LoRA, and then needs a simple interface so the rest of the team can run it without opening ComfyUI. A product team wants customers to be able to restyle an image, with the app sending each request to a workflow the team controls, scaling with traffic. An operations team wants new catalog items to run through a workflow automatically every week, without anyone watching the graph.

All three are variations on the same need: take a graph that encodes real expertise and let other people use it without exposing or rebuilding it. Comfy API targets exactly that. It lets you give your team a tool, add a feature to a product, or automate repeatable creative work, without any of those consumers needing to understand the graph.

For Team and Enterprise plans, teammates can work from the same Build, with the version, models, custom nodes and dependencies pinned together. Enterprise customers get Managed Builds and governance controls to standardize approved versions and dependencies across teams. That last capability addresses a real risk in large organizations, where three teams maintain three incompatible ComfyUI environments and nobody can tell which one produced a given asset.

The bigger picture

What ComfyUI has effectively done is upgrade itself from a local tool into a hosted media-generation stack. The engine remains open source, and Builds stay portable to hardware you own, so the company is not trying to trap the workflow inside its cloud. That portability is a notable choice in a market where hosted tools usually pull in the other direction.

The launch also says something about where the value in image and video generation has migrated. A year ago the differentiation was in the model. Now that capable models are available to everyone, and open weights put many of them on local hardware, the differentiation is shifting to the workflow: the specific sequence of nodes and settings that turns a general model into a reliable output for one narrow purpose. Those workflows are where domain expertise gets encoded, and until now they have been hard to operationalize.

Whether Comfy API becomes the standard path for that is an open question. It requires a paid plan, it runs on ComfyUI's own platform, and it does not compete on raw price with the specialized GPU clouds. What it does offer is the removal of the least creative and most error-prone part of shipping AI media. For a lot of small teams, that trade is the whole reason they can turn an internal tool into something they can sell.