Wan2GP API Server

The FastAPI server includes its own monitoring dashboard.

http://192.168.1.53:7861/monitor?token=YOUR_SECRET_TOKEN

http://192.168.1.53:7861/ui?token=YOUR_SECRET_TOKEN

- API status
- loaded model
- total jobs
- active jobs
- completed jobs
- failed jobs
- job status
- queue position
- generation mode
- progress
- current phase
- current step
- requester IP
- requester user-agent
- prompt excerpt
- input files
- LoRA information
- generation duration
- MP4 download link

Method 1: let your agent install the skill from GitHub

The simplest method is to give your AI agent the URL of this GitHub repository and ask it to install the skill itself.

Example instruction to give to your agent:

Install the Wan2GP video generation skill from this GitHub repository. Read the README, copy the agent skill files into your skill workspace, and make the skill available for use.

This part must be done on the Linux machines running the agents.

mkdir -p ~/.openclaw/workspace/skills/wan2gp_video cp wan2gp_video_agent_skill/* ~/.openclaw/workspace/skills/wan2gp_video/

export WAN2GP_URL="http://192.168.1.53:7861" export WAN2GP_TOKEN="YOUR_SECRET_TOKEN"

It is recommended to use environment variables instead of hardcoding the token in the Python file.

Add this to the system prompt or skill configuration of your agent:

You have access to a skill called Wan2GP Video. Use this skill whenever the user asks for video generation. Choose the mode automatically: - text only: t2v - image + prompt: i2v - start image + end image + prompt: i2v_end - audio + prompt: s2v - audio + image + prompt: s2v_i2v - audio + image + explicit LoRA request: s2v_i2v_lora After submitting the job, retrieve the job_id, monitor progress with get_job_status, wait until the job is complete, download the generated MP4, then return the video file to the user. When useful, provide the user with the built-in monitor URL so they can follow the queue visually.

from wan2gp_skill import generate_video result = generate_video( mode="t2v", prompt="A cinematic shot of a small robot walking under neon rain, realistic lighting", duration_seconds=3, output_path="/tmp/robot.mp4", verbose=True, ) print(result["saved_file"])

from wan2gp_skill import generate_video result = generate_video( mode="i2v", prompt="A cinematic close-up portrait, subtle natural movement, warm daylight", image_path="/tmp/reference.png", duration_seconds=3, output_path="/tmp/i2v.mp4", verbose=True, ) print(result["saved_file"])

from wan2gp_skill import generate_video result = generate_video( mode="i2v_end", prompt="The subject slowly turns toward the camera as the lighting shifts from warm daylight to blue evening light", image_start_path="/tmp/start.png", image_end_path="/tmp/end.png", duration_seconds=4, output_path="/tmp/i2v_end.mp4", verbose=True, ) print(result["saved_file"])

from wan2gp_skill import generate_video result = generate_video( mode="s2v", prompt="A cinematic dialogue scene with natural facial expression and perfect lip sync", audio_path="/tmp/voice.mp3", duration_seconds=6, output_path="/tmp/s2v.mp4", verbose=True, ) print(result["saved_file"])

from wan2gp_skill import generate_video result = generate_video( mode="s2v_i2v", prompt="The woman speaks naturally in French in front of the camera, soft studio lighting, realistic facial motion, perfect lip sync", image_path="/tmp/reference.png", audio_path="/tmp/voice.mp3", duration_seconds=6, output_path="/tmp/s2v_i2v.mp4", verbose=True, ) print(result["saved_file"])

from wan2gp_skill import generate_video result = generate_video( mode="s2v_i2v_lora", prompt="The woman speaks in French in front of the camera with perfect lip sync, calm studio ambiance, subtle cinematic camera movement", image_path="/tmp/reference.png", audio_path="/tmp/voice.mp3", duration_seconds=6, output_path="/tmp/result.mp4", verbose=True, ) print(result["saved_file"])

This is useful when several agents submit jobs and the queue is monitored from the built-in dashboard.

from wan2gp_skill import generate_video result = generate_video( mode="t2v", prompt="A futuristic hospital corridor, cinematic lighting, slow dolly forward", duration_seconds=4, wait=False, ) print(result["job_id"]) print(result["monitor_url"])

GET /health GET /model GET /jobs GET /jobs/{job_id} GET /download/{job_id}/{filename} GET /monitor?token=YOUR_SECRET_TOKEN GET /ui?token=YOUR_SECRET_TOKEN GET /monitor/download/{job_id}/{filename}?token=YOUR_SECRET_TOKEN POST /generate/t2v POST /generate/i2v POST /generate/i2v_end POST /generate/s2v POST /generate/s2v_i2v POST /generate/s2v_i2v_lora

Protected API endpoints require a Bearer token:

The browser monitoring endpoints accept the token as a query parameter:

Returns basic API status and available modes.

Returns the fixed model information, default generation settings, template file path, and supported mode controls.

Returns all jobs currently known by the API server.

Jobs are stored in memory. If the server restarts, the job history is cleared.

Downloads a generated MP4 using Bearer token authentication.

Displays the built-in HTML queue dashboard.

GET /monitor/download/{job_id}/{filename}

For LTX 2.3 video generation, write the prompt as a clear cinematic direction, not as a list of keywords.
- shot type
- camera movement
- environment
- lighting
- subject
- visible action
- mood expressed through physical details
- audio or dialogue when needed

The camera starts in a tight cinematic close-up, then slowly pushes forward as the woman raises her eyes toward the lens. Warm studio light reflects softly on her face. She breathes in, pauses, and says in French, "Je crois que j'ai enfin compris." Her voice is quiet and sincere. After the line, she gives a small uncertain smile while the background remains softly blurred.

For Image-to-Video, do not redescribe what is already visible in the reference image. Describe only:

- camera movement
- subject movement
- environmental changes
- lighting changes
- facial expression changes
- dialogue or sound timing

If the video contains dialogue, include the dialogue directly inside the prompt exactly where it happens in the scene.

The Wan2GP API server must be running before agents can generate videos.

The first generation after startup may be slower if the model has to be loaded into VRAM.

Jobs are stored in memory on the API server.