agno-multimodal

name	agno-multimodal
description	Build multimodal Agno agents that handle images, audio, and video. Covers image analysis, audio input/output, video captions, and file processing. Trigger this skill when: processing images with agents, handling audio or video, using vision capabilities, or asking "how do I build a multimodal agent?"
license	Apache-2.0
metadata	{"version":"1.0.0","author":"agno-team","tags":["multimodal","images","audio","video","vision","agno"]}

Build Multimodal Agno Agents

Agno agents can process images, audio, and video using model vision and multimodal capabilities. Install with pip install agno.

Image Analysis

From URLs

from agno.agent import Agent
from agno.media import Image

agent = Agent(
    model="openai:gpt-4o",
    markdown=True,
)

agent.print_response(
    "What's in this image?",
    images=[Image(url="https://example.com/photo.jpg")],
    stream=True,
)

From Local Files

from agno.agent import Agent
from agno.media import Image

agent = Agent(
    model="openai:gpt-4o",
    markdown=True,
)

agent.print_response(
    "Describe this image in detail.",
    images=[Image(filepath="path/to/image.png")],
    stream=True,
)

Multiple Images

agent.print_response(
    "Compare these two images.",
    images=[
        Image(url="https://example.com/before.jpg"),
        Image(url="https://example.com/after.jpg"),
    ],
    stream=True,
)

Audio Input

from agno.agent import Agent
from agno.media import Audio

agent = Agent(
    model="openai:gpt-4o-audio-preview",
    markdown=True,
)

agent.print_response(
    "Transcribe and summarize this audio.",
    audio=[Audio(filepath="path/to/recording.mp3")],
    stream=True,
)

Audio Output

Generate spoken responses:

from agno.agent import Agent
from agno.models.openai import OpenAIChat

agent = Agent(
    model=OpenAIChat(
        id="gpt-4o-audio-preview",
        modalities=["text", "audio"],
        audio={"voice": "alloy", "format": "wav"},
    ),
)

response = agent.run("Tell me a short story about a robot.")

# Save audio output
if response.audio:
    with open("output.wav", "wb") as f:
        f.write(response.audio)

Video Analysis

from agno.agent import Agent
from agno.media import Video

agent = Agent(
    model="google:gemini-2.0-flash",
    markdown=True,
)

agent.print_response(
    "Describe what happens in this video.",
    videos=[Video(filepath="path/to/video.mp4")],
    stream=True,
)

Image with Tools

Combine vision with tools for richer analysis:

from agno.agent import Agent
from agno.media import Image
from agno.tools.websearch import WebSearchTools

agent = Agent(
    model="openai:gpt-4o",
    tools=[WebSearchTools()],
    instructions=["Identify objects in images and search for more info."],
    markdown=True,
)

agent.print_response(
    "What landmark is this? Give me historical facts.",
    images=[Image(filepath="landmark.jpg")],
    stream=True,
)

Multimodal Agent with Structured Output

Extract structured data from images:

from pydantic import BaseModel, Field
from agno.agent import Agent
from agno.media import Image

class ReceiptData(BaseModel):
    store_name: str = Field(..., description="Name of the store")
    total: float = Field(..., description="Total amount")
    items: list[str] = Field(..., description="List of items purchased")

agent = Agent(
    model="openai:gpt-4o",
    output_schema=ReceiptData,
)

response = agent.run(
    "Extract the receipt data.",
    images=[Image(filepath="receipt.jpg")],
)

receipt: ReceiptData = response.content
print(f"Store: {receipt.store_name}, Total: ${receipt.total:.2f}")

Model Support for Multimodal

Capability	Models
Image input	GPT-4o, Claude Sonnet/Opus, Gemini
Audio input	GPT-4o-audio-preview
Audio output	GPT-4o-audio-preview
Video input	Gemini 2.0 Flash

Anti-Patterns

• Don't send huge images — resize before sending to save tokens and latency
• Don't use vision models for text-only tasks — they cost more per token
• Don't forget model compatibility — not all models support all modalities
• Don't mix audio and vision in one call unless the model supports it
• Don't skip filepath= for local files — use url= only for remote resources

Further Reading

For advanced multimodal patterns and provider-specific options, read references/api-patterns.md.