dkackman/diffusers-workflow repository overview

⁠diffusers-workflow

Your GPU, as something an agent can drive. This image runs the diffusers-workflow⁠ server: Hugging Face Diffusers image, video and audio generation as jobs, with a web UI and an MCP endpoint that Claude Code (or any MCP client) uses to pick a workflow, quote its cost, run it and look at the result.

⁠What's inside

Python, torch (CUDA 13), the engine, the web UI and the MCP server. diffusers lives on the data volume so it can follow GitHub main, where new model pipelines land first. It is installed on first start and updated only when you ask.

⁠Requirements

• Linux with an NVIDIA GPU, Turing or newer (RTX 20-series and up)
• NVIDIA driver 580 or newer
• NVIDIA Container Toolkit⁠

macOS is not supported: Docker on a Mac has no access to the GPU. Use the native install there.

⁠Quick start

docker run -d --name dw --gpus all --shm-size 16g \
  -p 127.0.0.1:8765:8765 -v dw-data:/data \
  dkackman/diffusers-workflow:latest

docker logs -f dw                                         # first start installs diffusers
docker exec dw python /opt/dw/entrypoint.py login-url     # prints the login link

Open the link to sign in to the web UI. The server always requires a token. It is generated on first start and never written to the log. Pass -e HF_TOKEN=... for gated models such as Flux, LTX and MiniMax.

Connect Claude Code:

claude mcp add --transport http dw http://<host>:8765/mcp \
  --header "Authorization: Bearer <token>"

To reach it from your LAN, publish on 0.0.0.0:8765:8765 instead.

⁠With a native install on the same box

The repo's docker/compose.yaml mounts the workspace, the Hugging Face cache and the compile caches at their host paths. A container and a native server can then share models, outputs and job history. Setup, diffusers pinning and options are in docs/DOCKER.md⁠.

⁠Tags

latest and one tag per release from 0.18.0 on.

⁠Links

Source and docs⁠ · Remote GPU server⁠ · MCP tools⁠