Gemini Video Generation

Use Veo 3.1 through GeminiTools on Vertex AI and decode the returned base64 content before saving the MP4.

Google discontinued the example's default veo-2.0-generate-001 endpoint after June 30, 2026 and recommends veo-3.1-generate-001. The source also converts video.content to the string representation of a bytes object, which corrupts the saved MP4. Apply the corrections below. See Vertex AI release notes.

gemini_video_generation.py

"""Example: Using the GeminiTools Toolkit for Video Generation

An Agent using the Gemini video generation tool.

Video generation only works with Vertex AI.
Make sure you have set the GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION environment variables.

Example prompts to try:
- "Generate a 5-second video of a kitten playing a piano"
- "Create a short looping animation of a neon city skyline at dusk"

Run `uv pip install google-genai agno` to install the necessary dependencies.
"""

from agno.agent import Agent
from agno.models.openai import OpenAIChat
from agno.tools.models.gemini import GeminiTools
from agno.utils.media import save_base64_data

# ---------------------------------------------------------------------------
# Create Agent
# ---------------------------------------------------------------------------


agent = Agent(
    model=OpenAIChat(id="gpt-5.6-luna"),
    tools=[GeminiTools(vertexai=True)],  # Video Generation only works on VertexAI mode
)

# ---------------------------------------------------------------------------
# Run Agent
# ---------------------------------------------------------------------------
if __name__ == "__main__":
    agent.print_response(
        "create a video of a cat driving at top speed",
    )
    response = agent.get_last_run_output()
    if response and response.videos:
        for video in response.videos:
            if video.content:
                save_base64_data(
                    base64_data=str(video.content),
                    output_path=f"tmp/cat_driving_{video.id}.mp4",
                )

Run the Example

Set up your virtual environment

uv venv --python 3.12
source .venv/bin/activate

Install dependencies

uv pip install -U agno google-genai openai

Export environment variables

export GOOGLE_CLOUD_LOCATION="your_google_cloud_location_here"
export GOOGLE_CLOUD_PROJECT="your_google_cloud_project_here"
export OPENAI_API_KEY="your_openai_api_key_here"

Authenticate with Google Cloud

gcloud auth application-default login

Enable Vertex AI access

Enable the Vertex AI API and billing in your project, grant the signed-in identity access to generate video, and choose a region supported by the Veo 3.1 model.

Use Veo 3.1

Replace GeminiTools(vertexai=True) with GeminiTools(vertexai=True, video_generation_model="veo-3.1-generate-001", enable_generate_image=False) in the saved file.

Capture the run result

Replace agent.print_response(...) and response = agent.get_last_run_output() with response = agent.run("Create a short video of a cat driving"). The saved example has no database or enabled session cache for the last-run lookup.

Decode the returned content

Replace base64_data=str(video.content) with base64_data=video.content.decode("utf-8") so save_base64_data() receives the base64 string instead of a bytes representation.

Run the example

Save the code above as gemini_video_generation.py, then run:

python gemini_video_generation.py

The toolkit exposes only a prompt, sets enhance_prompt=True, and polls until the operation finishes without a toolkit timeout. It does not expose duration or other Veo generation settings; the five-second source prompt is not a duration guarantee. Validate the selected model and SDK configuration in your project before relying on the generated result.

Full source: cookbook/91_tools/models/gemini_video_generation.py