AIMLAPITools gives agents image, video, speech and transcription

Agno's AIMLAPITools puts each media capability behind its own enable_* flag and its own model setting, so an agent gets only the tools you pick. Out of the box it uses openai/gpt-image-2 for images, bytedance/seedance-2-5 for video, openai/tts-1 for speech and deepgram/nova-3 for transcription, and every model can be swapped for another one on the gateway.

The toolkit reads the same AIMLAPI_API_KEY as the AIMLAPI chat model, so one key runs both:

from agno.agent import Agent
from agno.models.aimlapi import AIMLAPI
from agno.tools.models.aimlapi import AIMLAPITools
 
tools = AIMLAPITools(enable_generate_video=False, base_dir="tmp")
print([tool.__name__ for tool in tools.tools])
# ['generate_image', 'generate_speech', 'transcribe_audio']
 
agent = Agent(
    model=AIMLAPI(id="gpt-5.6-luna"),
    tools=[tools],
    send_media_to_model=False,
)
 
response = agent.run("Generate an image of a lighthouse in a storm")
for image in response.images or []:
    print(image.mime_type, len(image.content))

The image, video and speech tools download the finished file and return a ToolResult with an Agno Image, Video or Audio holding the bytes, the MIME type and, for images and video, the original prompt. transcribe_audio hands the transcript straight back to the model as text.

Video and transcription run as jobs on AI/ML API's side. The toolkit submits the job and polls it, every 5 seconds for up to 900 for video and every 2 seconds for up to 300 for a transcript, and it retries a transient poll failure up to three times so a network blip doesn't abandon a job that is still billing. Async agents get the same tools as coroutines under arun.

transcribe_audio takes an https URL or a path inside base_dir, and with restrict_to_base_dir=True it refuses anything outside that folder, which keeps a prompt from making the agent upload an arbitrary file. Asset downloads never carry your API key.

See the cookbook, and learn more about AI/ML API in the documentation.