langchain-wavespeed

LangChain integration for WaveSpeed AI — run state-of-the-art image and video generation models from your LangChain agents and chains.

Installation

pip install -U langchain-wavespeed

Set your API key (get one at wavespeed.ai):

export WAVESPEED_API_KEY="your-api-key"

Tools

WaveSpeedImageGeneration

Generate images from text prompts (defaults to bytedance/seedream-v5.0-pro):

from langchain_wavespeed import WaveSpeedImageGeneration

tool = WaveSpeedImageGeneration()
url = tool.invoke(
    {
        "prompt": "A red panda drinking boba tea, studio lighting",
        "resolution": "2k",      # optional: "1k" | "1.5k" | "2k"
        "aspect_ratio": "16:9",  # optional
    }
)
print(url)  # https://.../output.png

WaveSpeedVideoGeneration

Generate videos from text prompts (defaults to bytedance/seedance-2.5/text-to-video):

from langchain_wavespeed import WaveSpeedVideoGeneration

tool = WaveSpeedVideoGeneration()
url = tool.invoke({"prompt": "A drone shot over a glacier at sunrise", "duration": 5})

WaveSpeedRunModel

Run any model on the WaveSpeed platform by id (browse the catalog at wavespeed.ai/models):

from langchain_wavespeed import WaveSpeedRunModel

tool = WaveSpeedRunModel()
url = tool.invoke({
    "model": "wavespeed-ai/z-image/turbo",
    "input": {"prompt": "A lighthouse at dusk"},
})

Use with an agent

from langchain.agents import create_agent
from langchain_wavespeed import WaveSpeedImageGeneration, WaveSpeedVideoGeneration

agent = create_agent(
    "openai:gpt-5",
    tools=[WaveSpeedImageGeneration(), WaveSpeedVideoGeneration()],
)
result = agent.invoke(
    {"messages": [{"role": "user", "content": "Make me a picture of a corgi surfing."}]}
)

Configuration

All tools accept:

Parameter	Default	Description
api_key	WAVESPEED_API_KEY env var	WaveSpeed API key
model	tool-specific	Model id to run (image/video tools)
timeout	600.0	Max seconds to wait for a prediction (None waits forever)
poll_interval	2.0	Seconds between result polls

When a prediction fails or times out, the tool raises ToolException with the platform's error text and the task id, so a paid task stays traceable (and an agent can read the failure instead of crashing the run). A timeout only stops the waiting - the task keeps running server-side.

await tool.ainvoke(...) works, but note that the underlying WaveSpeed SDK is synchronous: LangChain runs the blocking call in a worker thread, so it will not block your event loop, but it is not natively async I/O.

License

MIT

---

WaveSpeed AI — AI image & video generation platform. Try it in the browser: Image generator · Video generator