Workflow1111 Rebuilds AUTOMATIC1111 as MCP Tools

On 10 September 2026 the Gradio team published Workflow1111, a rebuild of most of AUTOMATIC1111's stable-diffusion-webui feature set as a single node canvas: eleven media pipelines, seventy-three nodes, one browser tab. It runs text-to-image, hi-res fix, image-to-image, prompt-matrix grids, VLM interrogation, detection-to-inpaint masks, ControlNet-style annotators, background removal, PNG Info and image-to-video. Nine of its outputs are REST endpoints, and the same nine are MCP tools an agent can call.

The interesting part is not the feature list. It is that the project it rebuilds has not shipped a tagged release in nineteen months, and that this replacement runs on a free CPU box.

What is on the canvas

Workflow1111 is built on gr.Workflow, the graph primitive Gradio shipped on 25 August. We covered the primitive itself at launch; this is the first serious application built on it. Everything on the canvas is one of four operator kinds: fn is a plain Python function, model calls a checkpoint through Inference Providers, space calls another Gradio Space on the Hub, and dataset pulls a row from a Hub dataset.

Two numbers from the post are worth keeping straight, because summaries will blur them. The graph has 73 nodes in total. Of those, 36 are operator nodes, 32 of the 36 are fn nodes, and 22 of those run entirely in-process with no network call at all. The team's phrasing is that "roughly two-thirds of the canvas keeps working if you lose your connection." The five ControlNet-style annotators (Canny, line art, sketch, luma-depth, posterize) are pure NumPy with no model behind them, at about half a second each on CPU.

The pipelines that do reach out reach a long way. Hi-res fix and image-to-image are the same FLUX.1-Kontext node with different instructions. Prompt enhancement runs Qwen3-4B and caps the result at forty tags. Interrogate uses Qwen2.5-VL plus a ViT classifier that returned "restaurant 51.9%, tobacco shop 15.6%, toyshop 9.1%" on the demo image, with both nodes running in parallel because they sit at the same dependency depth. Inpaint masks come from DETR object detection instead of hand-painting. Image-to-video is Wan 2.2 I2V A14B. Upscaling and background removal are space nodes pointed at AuraSR and BRIA RMBG-2.0.

One detail the announcement does not draw attention to: the checkpoints the Space declares are FLUX.1-dev, FLUX.1-Kontext-dev and FLUX.1-schnell. This is an AUTOMATIC1111 rebuild that does not run Stable Diffusion out of the box.

AUTOMATIC1111 has not shipped a release since February 2025

This is the fact that makes the launch worth more than a changelog entry, and it is checkable in ten seconds. The stable-diffusion-webui repository's most recent tagged release is v1.10.1, published on 9 February 2025. That is nineteen months ago. The last push to the repository was 2 March 2026, six months ago. It carries 2,509 open issues.

And it is still, by a wide margin, the most-starred image generation interface in open source: 164,885 stars against ComfyUI's 132,352. A frozen project holds a 32,000-star lead over the one that shipped v0.35.0 yesterday. That gap is the shape of the problem. An enormous number of people learned image generation through the A1111 tab layout, the workflow muscle memory is still in their hands, and the software under it stopped moving a year and a half ago.

Every credible answer so far has asked those people to learn a graph. ComfyUI, which we covered when it added Seedance, FishAudio and Flux nodes in August, is a graph. Workflow1111 is also a graph, so it does not solve the retraining problem either. What it does differently is keep the A1111 feature inventory as the organising principle, tab by tab, rather than starting from what the node system can express.

Workflow1111 vs ComfyUI vs AUTOMATIC1111

Read the last two rows together with the GPU row. Workflow1111 wins on access and loses on control. ComfyUI and A1111 win on control and demand a GPU and an install.

The whole thing runs on a CPU Space

The Space's public runtime metadata reports hardware cpu-basic. Not a shared GPU, not ZeroGPU: a free CPU container. Every diffusion call, every VLM call, every video generation leaves the box and runs on Inference Providers, billed against the quota of whoever signed in. The Space itself holds no credentials.

That inverts the economics of the thing it replaces. Running A1111 meant owning or renting a GPU and paying for it whether or not you generated anything. Running Workflow1111 means signing in with a Hugging Face account, using your own quota, and paying for exactly the calls you make. For a creator evaluating whether a pipeline is worth building at all, that is the difference between a hardware decision and a five-minute test.

It also inverts who bears the risk. Your pipeline now depends on Inference Providers being up, on the AuraSR and BRIA Spaces staying published, and on the FLUX checkpoints remaining served. A space node is a dependency on somebody else's deployment, and unlike a Python package it cannot be pinned to a version you control. The 22 in-process nodes are the part nobody can take away from you.

The escape hatch is real, though, and it is the same mechanism. Because an fn node is just Python, it can load a checkpoint and run it on hardware you do own. The team demonstrates this with a companion Space running FastH3, the four-step distillation of MiniMax-H3, on ZeroGPU behind a single @spaces.GPU decorator. Point bind= at a function that loads a local checkpoint, call .launch() on your own machine, and the same canvas drives your own GPU.

Nine endpoints that double as MCP tools

Every output node becomes a REST endpoint with no routing written by hand. Workflow1111 exposes nine: /image, /edited_image, /generated_prompt, /recovered_prompt, /detected_objects, /x_y_grid, /upscaled_local, /annotator_map and /png_info.

Launch with mcp_server=True, documented in Gradio's MCP guide, and those same nine become tools in any MCP client. The config is four lines pointed at https://ysharma-workflow1111.hf.space/gradio_api/mcp/ with an X-HF-Token header. Each caller sends their own token, so a shared server never accumulates other people's credentials.

This is the part that has no equivalent in either comparison column. An agent in Claude Code or Cursor can now generate an image, read a prompt back out of a PNG, run object detection and build an inpaint mask as ordinary steps inside a larger task, with no glue code and no custom tool definitions. The image pipeline stops being an app a human operates and becomes a set of functions a program calls. Whether that is what you want is a separate question, but it is a capability neither A1111 nor a stock ComfyUI install offers.

The licence line that will matter commercially

Three projects, three different licences, and the difference is not cosmetic. AUTOMATIC1111's webui is AGPL-3.0, the strictest of the three: run a modified version as a network service and you owe your users the source. ComfyUI is GPL-3.0. The Workflow1111 Space declares MIT, and Gradio itself is Apache-2.0.

For anyone building a product on top of an image pipeline rather than just using one, that ordering is the whole decision. A studio that wants to embed a generation canvas inside a closed client tool can fork an MIT Space and ship it. The same studio cannot do that with an AGPL codebase without opening its own. This has been a quiet drag on A1111-derived commercial work for years, and it is one of the few dimensions on which the newcomer is not making a trade-off.

What it does not do

Three limits are worth stating plainly, because the launch post is honest about them and the coverage will not be.

There is no loop operator. The prompt-matrix pipeline combines one base prompt with four suffixes, and the way it does that is four separate text-to-image nodes sitting side by side on the canvas. They run in parallel, which is fast, but a five-variant matrix means editing the graph. Anyone whose A1111 habit is sweeping a parameter across twenty values does not have a path here yet.

There is no local checkpoint by default. The model_id field selects from what Inference Providers serves. Custom LoRAs, community fine-tunes and the entire Civitai-shaped half of the A1111 ecosystem are not addressed by this release at all, and for a large share of A1111 users that ecosystem was the reason to use A1111.

And it is a demonstration, not a maintained product. The Space was created on 17 August and last modified on 8 September. It ships test_api.py, test_nodes.py and test_pipelines.py, which is more rigour than most Spaces get, but the invitation in the post is explicitly to duplicate it and rewire it rather than to adopt it as-is.

What to do with this

If you are still on a frozen A1111 install, open the Space, sign in with a Hugging Face token and run the interrogate and detection-to-mask pipelines against your own images. Those two are the pipelines where the rebuild is genuinely better than the original rather than merely equivalent, because a VLM writing a prompt beats CLIP interrogate and a detector drawing a mask beats painting one by hand. That test costs a few minutes of quota and tells you whether the graph model fits your hands.

If you build tools, the more interesting move is to ignore the canvas and wire the MCP endpoint into your agent instead. Nine typed tools covering generation, editing, interrogation, detection and metadata is a usable image toolbelt, and standing up an equivalent yourself is a weekend of work you no longer have to do.

If you depend on custom checkpoints and LoRAs, wait. Nothing in this release touches that, and ComfyUI remains the answer.

Frequently asked questions

Is Workflow1111 a fork of AUTOMATIC1111?

No. It shares no code with stable-diffusion-webui. It is a rebuild of the feature set as a Gradio workflow graph, written by the Gradio team as a demonstration of what gr.Workflow can express.

Do I need a GPU to run it?

No. The hosted Space runs on cpu-basic hardware and routes every model call to Inference Providers under your own signed-in quota. If you want local execution, an fn node can load a checkpoint and run on your own hardware.

Is AUTOMATIC1111 dead?

The repository is not archived, but the most recent tagged release is v1.10.1 from 9 February 2025 and the last push was 2 March 2026. It has 2,509 open issues. It still has more GitHub stars than ComfyUI, so it is widely installed rather than actively developed.

Can Claude or Cursor drive it?

Yes. Launching with mcp_server=True exposes all nine output endpoints as MCP tools. Point any MCP client at the server URL with an X-HF-Token header and the pipelines become callable tools inside a larger agent task.

Does it support LoRAs and custom checkpoints?

Not in this release. Model selection happens through a model_id field backed by Inference Providers, and the declared checkpoints are FLUX.1-dev, FLUX.1-Kontext-dev and FLUX.1-schnell. Custom fine-tunes remain a ComfyUI or local-install job.

What licence is it under?

The Space declares MIT and Gradio is Apache-2.0, which is more permissive than ComfyUI's GPL-3.0 and considerably more permissive than AUTOMATIC1111's AGPL-3.0. That matters if you intend to embed a pipeline in a commercial product.

How many nodes does it actually have?

73 nodes across eleven pipelines. Of those, 36 are operator nodes, 32 of the 36 are Python function nodes, and 22 run in-process with no network call.