AI & Machine Learningreplicate.com &nearr;

Replicate API for AI Agents

Run open-source ML models via API

Replicate lets you run open-source machine learning models through a simple API. AI agents can use Replicate to generate images, transcribe audio, process video, and run any model from the Replicate catalog.

What AI agents can do with Replicate

Structured actions an AI agent can execute through the Replicate API

Action

Description

Inputs

Outputs

createPrediction

Run a model prediction

model, input, webhook

prediction_id, status, output

getPrediction

Check prediction status

prediction_id

status, output, metrics

listModels

Browse available models

query, cursor

models[], next_cursor

Use cases for Replicate + AI agents

Image generation with Stable Diffusion and FLUX
Audio transcription with Whisper
Video processing and generation
Running custom fine-tuned models
Multi-model AI pipelines

How to connect Replicate to an AI agent

1Get your Replicate API token
2Generate an AgentSpec for model prediction actions
3Define actions for each model you want agents to use
4Publish for discovery
5Handle async predictions with polling or webhooks

Best practices

✓Use webhooks for long-running predictions instead of polling

✓Include model version in predictions for reproducibility

✓Cache model outputs when inputs are identical

✓Add reasoning docs about which model to use for which task

✓Handle cold start latency in your agent workflow

Frequently asked questions

What models can AI agents run on Replicate?+

Any model in the Replicate catalog — image generators (FLUX, SDXL), language models (Llama, Mistral), audio models (Whisper), and thousands more. Agents can select models based on task requirements.

How do agents handle async predictions?+

Create a prediction, then poll with getPrediction until status is "succeeded". For better performance, use webhooks to get notified when predictions complete.

Is Replicate cost-effective for AI agents?+

You pay per second of compute, with no minimum. Cold starts are free. This makes it efficient for agents that need occasional access to powerful models without maintaining infrastructure.