For Agents
Train, deploy, and run predictions on custom AutoML models for vision, language, translation, and tabular tasks in Google Cloud.
Get started with Cloud AutoML API in minutes using your preferred integration method.
# Add to your MCP client config (Claude Desktop, Cursor, Windsurf)
{
"jentic": {
"url": "https://api.jentic.com/mcp",
"auth": "oauth"
}
}
# Then ask your agent:
"run a batch prediction with a Google AutoML model"
# → Jentic returns the GET /events tool with parameter schema, agent executes.What an agent can do with Cloud AutoML API API.
Create and import datasets for image, text, translation, and tabular AutoML tasks
Train a custom model from a labelled dataset and track progress via the operations endpoint
Deploy and undeploy a trained model with /v1beta1/{+name}:deploy and :undeploy
Run batch predictions on Cloud Storage inputs via /v1beta1/{+name}:batchPredict
GET STARTED
Use for: I need to train a custom image classifier on labelled product photos, Run a batch prediction on 50,000 images stored in Cloud Storage, Deploy a previously trained AutoML model so it can serve online predictions, List all AutoML models in a project and their deployment state
Not supported: Does not generate text, run general LLM inference, or replace Vertex AI for new training workloads — use for AutoML datasets, models, and predictions on existing AutoML resources only.
The Cloud AutoML API trains and serves custom machine learning models for image classification, object detection, text classification, translation, and tabular prediction without requiring the caller to write training code. It exposes operations to manage datasets, kick off training, deploy or undeploy models, and run online or batch predictions. Long-running operations track training and deployment progress, and the API integrates with Cloud Storage for dataset import and batch prediction results. Note that AutoML is in maintenance with Vertex AI as Google's go-forward platform.
Cancel long-running operations with /v1beta1/{+name}:cancel when training is no longer needed
List locations available to a project for AutoML resources before creating a dataset
Patterns agents use Cloud AutoML API API for, with concrete tasks.
★ Custom Image Classification Model Training
Train a vision model on a labelled image dataset stored in Cloud Storage by creating a dataset, importing labels, and starting a training operation. The API returns a long-running operation handle to poll, and once training completes the model can be deployed and queried for online predictions. Suitable when an off-the-shelf vision API does not match the domain.
Create an image dataset, POST a training request to /v1beta1/{parent}/models, and poll the operation until done.
Batch Prediction on Stored Datasets
Score a large batch of inputs in one call by pointing AutoML at Cloud Storage URIs and running batchPredict against a deployed model. Results are written back to Cloud Storage, which keeps the calling system simple and avoids paginating thousands of online prediction calls.
POST a batch predict request with input GCS URIs and an output GCS prefix to /v1beta1/{name}:batchPredict, then watch the operation.
Model Deployment and Cost Control
Deploy AutoML models only when traffic is needed and undeploy them when idle to avoid serving charges. The API exposes deploy and undeploy as discrete long-running operations, making it easy to drive cost-control workflows from a scheduler or an event in production.
POST to /v1beta1/{name}:undeploy at the end of business hours and to :deploy at the start of the next day.
AI Agent Triage with Custom Vision Models via Jentic
Use an AI agent to route incoming images through a domain-specific AutoML classifier before applying business logic. Through Jentic the agent finds the predict and batchPredict operations by intent and runs them with project credentials supplied at execute time, so the model trained once is reused across many agent workflows.
Use Jentic search for 'predict with a Google AutoML model', load the predict schema, and call it with the image bytes for incoming items.
25 endpoints — the cloud automl api trains and serves custom machine learning models for image classification, object detection, text classification, translation, and tabular prediction without requiring the caller to write training code.
METHOD
PATH
DESCRIPTION
/v1beta1/{+name}
Get a dataset, model, or evaluation by resource name
/v1beta1/{+name}
Update mutable fields on a dataset or model
/v1beta1/{+name}
Delete a dataset, model, or operation
/v1beta1/{+name}:deploy
Deploy a trained model for online prediction
/v1beta1/{+name}:batchPredict
Run batch prediction on stored inputs
/v1beta1/{+name}:cancel
Cancel a long-running operation
/v1beta1/{+name}/operations
List operations for a parent resource
/v1beta1/{+name}
Get a dataset, model, or evaluation by resource name
/v1beta1/{+name}
Update mutable fields on a dataset or model
/v1beta1/{+name}
Delete a dataset, model, or operation
/v1beta1/{+name}:deploy
Deploy a trained model for online prediction
/v1beta1/{+name}:batchPredict
Run batch prediction on stored inputs
Three things that make agents converge on Jentic-routed access.
Credential isolation
Service account credentials with AutoML scopes are stored encrypted in the Jentic vault. Agents never receive the raw JSON key; each call mints a scoped, short-lived access token.
Intent-based discovery
Agents search Jentic with intents like 'train an image classifier' or 'run a batch prediction' and Jentic returns the matching AutoML operation with its input schema, so the agent does not need to navigate the discovery document.
Time to first call
Direct integration: 1-3 days for IAM setup, dataset import, and operation polling. Through Jentic: under 1 hour once the credential is in the vault.
Alternatives and complements available in the Jentic catalogue.
Vertex AI API
Google's go-forward unified ML platform that supersedes AutoML for new projects
Choose Vertex AI for new training and serving work; use AutoML only for existing models still on this surface.
AI Platform Training and Prediction API
Earlier custom-training and prediction surface in Google Cloud
Use only for legacy custom-trained models still hosted on AI Platform.
Cloud Vision API
Pretrained vision API covering general detection and labelling tasks
Use Cloud Vision when a pretrained model already covers the task; reach for AutoML only when the domain is custom.
Specific to using Cloud AutoML API API through Jentic.
What authentication does the Cloud AutoML API use?
The API uses Google OAuth 2.0 with cloud-platform scope, typically via a service account holding the roles/automl.* IAM role. Through Jentic the service account key is encrypted in the vault and the agent only sees a short-lived access token at execute time.
Can I run batch predictions with the Cloud AutoML API?
Yes. POST a request with input and output Cloud Storage URIs to /v1beta1/{name}:batchPredict against a deployed model. The call returns a long-running operation handle and writes prediction results to the specified output prefix.
How do I deploy or undeploy a model through Jentic?
Search Jentic for 'deploy a Google AutoML model' or 'undeploy a Google AutoML model'. Jentic returns the matching operation; execute it with the model resource name and Jentic supplies the credential from the vault.
What are the rate limits for the Cloud AutoML API?
Numeric limits are not declared in the spec; quotas are enforced per project on automl.googleapis.com. Check the project quota dashboard for prediction QPS and concurrent training jobs before scaling.
Is AutoML still actively developed?
Google now positions Vertex AI as the go-forward platform for custom ML, and AutoML is in maintenance for many product surfaces. New projects should evaluate Vertex AI for training and serving while still reachable via this API for in-flight workloads.
Can I cancel a training run that is no longer needed?
Yes. POST to /v1beta1/{name}:cancel on the operation resource for the training job. The operation transitions to cancelled and stops accruing training cost.
/v1beta1/{+name}:cancel
Cancel a long-running operation
/v1beta1/{+name}/operations
List operations for a parent resource