Wan2GP runs Wan, LTX-2, MiniMax H3, Flux, Qwen and other models. NodeTool talks to its MCP server, and the models run in Wan2GP’s own process and Python environment. Two parts use it:
- the
lib.wan2gpnodes, for workflows; - the
wangpprovider, which lists Wan2GP’s models in the chat composer, the Create view and ejected preset workflows.
| Node | What it does |
|---|---|
lib.wan2gp.TextToVideo |
Generates a clip from a prompt |
lib.wan2gp.ImageToVideo |
Animates a still image |
lib.wan2gp.Generate |
Runs any Wan2GP model with a settings dict on top of that model’s defaults. Takes start and end images, reference images, control, mask and source videos, audio, and a custom file, and returns an image, video, audio or 3D model |
lib.wan2gp.ListModels |
Lists the models your server offers and the media each one accepts |
Runs: locally only (the cloud profile excludes the nodes and the provider). Needs: a Wan2GP install or a Wan2GP server you run, and no API key.
WAN2GP_MCP_TOKENis needed only for a server behind a bearer token. Hardware: a supported GPU (table below) and about 12 GB of disk before any model.
1. Install Wan2GP
- Desktop app: open Tools → Package Manager → Engines & tools and install Wan2GP, or pick a Wan2GP model in the model menu and install it there. It needs the Python runtime, which installs first.
- CLI:
nodetool runtime add wan2gp. It needs a Python 3.10 or newer and uv, from the NodeTool conda environment or PATH.--pythonand--uvname others, and--profile RTX_40picks a hardware profile instead of the detected one.
The install downloads Wan2GP v19.02 from GitHub and runs Wan2GP’s own
unattended installer (setup.py install --env uv --auto). That installer
reads the GPU and picks the Python, PyTorch and attention kernels for it:
| Hardware | Python and PyTorch |
|---|---|
| NVIDIA RTX 20xx to 50xx, Windows or Linux | Python 3.11, PyTorch 2.10 with CUDA 13.0 |
| NVIDIA GTX 10xx, Windows or Linux | Python 3.10, PyTorch 2.7.1 with CUDA 12.8 |
| AMD, Windows or Linux | Python 3.12, PyTorch 2.13 with ROCm 10 |
| Apple Silicon | Python 3.11, PyTorch with MPS |
Without a supported graphics card the install stops with a message that
names the card it found and the profiles to choose from. A missing driver is
the usual cause. Allow about 12 GB before any model. Each model’s weights
download the first time it runs, into wan2gp/ckpts, which an update keeps.
The install ends by writing the model catalog, catalog.json beside the
release. It holds every model, the media it accepts, its voices or languages,
and whether its weights are on disk. The model menus read this file, so they
list Wan2GP’s models without starting Wan2GP. An install made before the
catalog existed writes it the first time a menu opens. A Package Manager
install or uninstall reloads the model menus that are open.
Download models
With Wan2GP installed, the model menu marks each Wan2GP model as ready or to
download, from the same file check Wan2GP’s own model selector uses. Download
fetches the model’s weights with Wan2GP’s own downloader into wan2gp/ckpts.
Before it starts, Hugging Face reports the size of the files still missing,
so the bar shows how much is left. Wan2GP sometimes fetches small files the
estimate leaves out, and past the estimate the bar shows the bytes written.
A cancelled download starts the cut-off file again. A generation also
downloads what its model lacks before it runs. A server you run yourself
manages its own files.
The server
NodeTool starts the installed server the first time a node or the provider
needs it, on a free port on 127.0.0.1. A finished model download starts it
too, so the first generation does not wait for Wan2GP to start. Opening a
model menu does not start it. It stops the server after 15 minutes
without calls, which frees the GPU memory Wan2GP holds, and when NodeTool
exits.
Run Wan2GP yourself instead
Follow Wan2GP’s installation guide in its own environment, then start its MCP server:
python wgp.py --mcp --mcp-async --mcp-transport streamable-http --mcp-host 127.0.0.1 --mcp-port 7866
The MCP endpoint is then http://127.0.0.1:7866/mcp. Bind to 0.0.0.0 only
on a trusted network or behind an authenticated reverse proxy. NodeTool needs
MCP API v2 or later, which has the wangp_session tool. It refuses an older
server with a message that says so.
2. Point NodeTool at the server
The nodes pick the server when they run, in this order:
- The node’s
server_urlfield, when it is not blank. WAN2GP_MCP_URL, set in Settings → Integrations or in the environment.- The Wan2GP the Package Manager installed, started on demand.
http://127.0.0.1:7866/mcp, for a server you run.
The wangp provider uses the same order without the first step. With
nothing installed and nothing answering at the default address, a generation
fails with an error that offers the install. A changed setting applies to the
next call, without a restart.
WAN2GP_MCP_TOKEN is a bearer token for a server behind an authenticating
proxy, such as the combined GPU image below. The nodes and the provider send
it on every request when it is set. A server that refuses the request names
the setting in its error.
Use presets with your server
The nodes take a plain prompt. In the chat composer, the Create view and
ejected preset workflows, pick a model under Wan2GP. The wangp provider
lists the server’s video and image models and the tasks each one supports, and
presets compile for the chosen model before
the prompt reaches the server. The list is empty while no server answers.
| Task | What the provider sends |
|---|---|
| Text to video | The prompt and the generation settings |
| Image to video | The first frame as image_start, and the last frame as image_end when given |
| Reference to video | Reference images as image_refs. Reference videos and audio are refused: use the Generate node for those |
| Text to image | The prompt with image_mode 1, which picks the image output on models that also make video |
| Image to image | The image as image_start or as image_refs, whichever role the model declares |
| Text to speech | The text as prompt, the voice (Qwen3 Custom Voice) or language as model_mode, the instructions or reference transcript as alt_prompt, and a reference voice as audio_guide on the models that take one |
| Text to music | The lyrics as prompt and the style as alt_prompt, or [Instrumental] without lyrics. Stable Audio takes the style as prompt. The duration as duration_seconds |
Speech models list their preset voices or languages. Music models that need a file, such as YuE2’s score, run only through the Generate node. The transcription models are not offered.
A resolution tier such as 720p sets the short side, and the aspect ratio,
16:9 by default, sets the long one. Wan2GP aligns the size to its own block.
Combined GPU worker
The nodetool-wan2gp
repository builds Dockerfile.combined, a GPU image that runs a pinned Wan2GP
release with its MCP server behind a proxy that requires a bearer token. Point
NodeTool at it with the token the container was started with:
export WAN2GP_MCP_URL=http://<gpu-host>:7866/mcp
export WAN2GP_MCP_TOKEN=<the container's WAN2GP_MCP_TOKEN>
Build and run instructions are in the nodetool-wan2gp README.
Licensing
Wan2GP is under the WanGP Community License 2.0. You may use it for free, including for client work. Selling it, white-labelling it, embedding it in a paid product, or offering paid hosted access to it needs a separate license from its author. NodeTool does not ship Wan2GP: the Package Manager downloads it from its GitHub repository when you install it. Each model has a separate license. MiniMax H3’s license excludes the European Union, the United Kingdom, South Korea and the United States, and LTX-2’s requires a paid license above 10 million US dollars in annual revenue. The nodetool-wan2gp README summarizes both. The combined image is for free, non-monetized use only. Read its NOTICE before you redistribute or sell access to it.
Related
- Supported Models: local engines and hardware
- Configuration:
WAN2GP_MCP_URLandWAN2GP_MCP_TOKEN - Node reference: the
lib.wan2gpnodes