Skip to main content

​
Replicate

Shanone’s Replicate integration gives your agent 31 tools for running and managing hosted AI models on Replicate — predictions, fine-tuning/training, deployments, and browsing the model catalog.

​
Getting Started

1

Create a Replicate API token

Generate a token from the Replicate account settings.
2

Add the token in Shanone

Go to Integrations in the Shanone dashboard, select Replicate, and paste in the token (r8_...).
3

Retry your request

Once saved, shanone_execute_tool calls for replicate_* tools will succeed.

​
Available Tools

Shanone provides 31 tools for Replicate, organized into these categories:
replicate_create_model, replicate_get_model, replicate_list_models, replicate_update_model, replicate_delete_model, replicate_search, replicate_list_model_examples, replicate_run_model, replicate_get_model_readme
replicate_create_prediction, replicate_get_prediction, replicate_list_predictions, replicate_cancel_prediction
replicate_create_training, replicate_get_training, replicate_list_trainings, replicate_cancel_training
replicate_create_deployment, replicate_get_deployment, replicate_list_deployments, replicate_update_deployment, replicate_delete_deployment, replicate_run_deployment
replicate_get_model_version, replicate_list_model_versions, replicate_delete_model_version, replicate_get_collection, replicate_list_collections
replicate_get_account, replicate_list_hardware, replicate_get_webhook_signing_secret

​
Common Use Cases

Model discovery

Search Replicate’s catalog for a model matching a task, then check its README for input parameters

On-demand inference

Run a model prediction directly, or against a pinned deployment for stable production behavior

Fine-tuning

Kick off a training job against a base model with a custom dataset and monitor progress

Webhook-driven pipelines

Verify incoming Replicate webhooks with the signing secret before processing prediction results

​
Troubleshooting

Predictions run asynchronously — poll replicate_get_prediction with the returned ID until status becomes succeeded or failed rather than expecting output in the initial response.
Each model defines its own input schema; check replicate_get_model_readme or the model’s version details for the exact expected parameters before calling replicate_run_model.
Training duration depends heavily on dataset size and base model — use replicate_get_training to check status rather than assuming a fixed runtime, and replicate_cancel_training if it needs to be stopped.