# diffusers-workflow MCP server

> Runs Hugging Face Diffusers image, video and audio generation workflows as jobs on your own GPU, driven by an AI agent.

- Repository: https://github.com/dkackman/diffusers-workflow
- Page: https://hostedmcp.org/servers/dkackman/diffusers-workflow
- Transport: —
- Stars: 18
- License: Apache-2.0

## FAQ

### How do I connect the diffusers-workflow MCP server to Claude Code?

Register the MCP server using the absolute path to dw-mcp in your virtual environment, e.g. claude mcp add dw -- "$(pwd)/venv/bin/dw-mcp". Make sure to use an absolute path, as relative paths are the most common setup mistake.

### Can I run diffusers-workflow on a remote GPU or in Docker?

Yes. Start the engine on the GPU machine with --mcp and connect over HTTP; nothing needs installing on the client. On Linux with an NVIDIA GPU you can also run the Docker image.

### Why did my run fail with a 401/403 from Hugging Face?

Most shipped models are gated. Request access on the model's Hugging Face page and run hf auth login once. The templates/text-to-image workflow uses a small ungated model, so it works as a first run.

### Can an agent estimate GPU cost before starting a job?

Yes. validate_workflow checks the workflow against the schema and pipeline signatures without loading models and quotes an estimate in GPU minutes. Work that spends GPU time or disk requires the agent to pass acknowledged_cost first.

### What is an MCP server for diffusers?

It lets an MCP client like Claude Code create, validate and run Diffusers image, video and audio generation workflows on your GPU, while you review and edit results in a browser UI.
