Manages Agent Platform serving endpoints. Use when you need to create, list, describe, update, or delete serving endpoints for model deployment on Agent Platform. Also use when troubleshooting endpoint permission, quota, or resource busy errors. Don't use for deploying models to endpoints or for running model evaluations.
A procedural skill for managing Agent Platform serving endpoints, the logical serving hosts that provide a stable URL for online predictions, using gcloud commands. It covers the endpoint lifecycle: listing, describing, creating, updating, and deleting endpoints, plus managing traffic splits between models deployed on the same endpoint. It is explicitly scoped to endpoint management and not to deploying models onto endpoints or running model evaluations, and it notes that an endpoint must exist before a model can be deployed to it.
A defining feature is its safety and confirmation model, which sorts actions into tiers. Tier R (read-only: list, describe, get) runs immediately with no confirmation. Tier M (mutating but reversible: create, update, traffic splitting) requires interactive Yes/No confirmation whose prompt must show the exact literal command string with all flags, and must never execute in the same turn as the prompt. Tier D (destructive and irreversible: delete) requires explicit typed confirmation requested immediately, before any pre-flight checks like describe, and likewise must wait for a new turn before executing. Environment setup is treated as mandatory: authenticate with gcloud and configure Application Default Credentials, set the active project, and always pass an explicit --region (never 'global'), asking the user for the region if it is not provided.
Use it when creating, inspecting, updating, or removing serving endpoints, or when troubleshooting endpoint errors. Each command is shown as a concrete gcloud ai endpoints invocation, and a troubleshooting section maps common failures to fixes: 403 permission denied (ensure the aiplatform.admin or owner role), quota exceeded (check the region's endpoint quota), and resource busy (models may still be undeploying). It also warns that all models must be undeployed before an endpoint can be deleted, and points to a related deploy skill for deploying and undeploying models.
Mutating actions (create, update, traffic splitting) require interactive Yes/No confirmation showing the literal command, and deletion requires explicit typed confirmation; read-only actions like list and describe need none.
No. Deletion needs typed confirmation requested up front, and the command must not run in the same turn; you must reply in a new turn before it executes.
Authenticate with gcloud (including Application Default Credentials), set the active project, and always specify an explicit --region rather than 'global'.
All models must be undeployed from the endpoint before it can be deleted; a resource-busy error can mean models are still being undeployed.
No. It is limited to endpoint management (create, list, describe, update, delete, traffic splits); deploying and undeploying models is handled by the related agent-platform-deploy skill.
Quick Setup:
.claude/skills/Repository
google/skills