Guide · Agent skill
How to use the Agent Platform Model Deploy agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The agent-platform-deploy skill is a set of instructions and helper scripts that lets an AI agent deploy open models from Google Cloud's Model Garden to hosted endpoints, check deployments that are still running, estimate hourly cost, troubleshoot errors, and clean up by undeploying models or deleting endpoints. It is aimed at engineers who already have a Google Cloud project with billing enabled. After this guide you will know how to install the skill, phrase requests so the agent picks it up, and how the approval steps protect you from accidental spending.
Agent Platform Model Deploygoogle/skills · Apache-2.0 Download the skill Quick start
- 1Download and unzip the skill. Download the ZIP from this page and unzip it so the skill folder with SKILL.md is in ~/.claude/skills/ or your project's .claude/skills/ folder.
- 2Prepare Google Cloud. Install the gcloud CLI, sign in, set application-default credentials and select your project with billing and the Agent Platform APIs enabled.
- 3Ask in plain language. Describe the model, project and region you want, for example a deploy or a cost estimate request.
- 4Review the confirmation card. Check the exact gcloud command, destination project, region, machine type, hourly cost and endpoint name before approving.
- 5Approve or stop. Reply with explicit approval to run a deploy or undeploy. The agent waits for your reply before executing.
- 6Check the result. Ask for the status of the operation, and read the operation name, endpoint ID or error message the agent reports.
- 7Clean up. Undeploy the model or delete the endpoint when you no longer need it, typing the confirmation the agent requests for deletes.
What the skill makes the agent do
An agent skill is a folder of instructions that an AI agent loads when a request matches its description. This one covers the deployment lifecycle for Model Garden models on Agent Platform endpoints. It runs gcloud commands for you, including listing the model catalog, checking whether a model is deployable, deploying, checking an in-progress deployment operation, and undeploying or deleting resources.
It also handles copying a first-party (1P) tuned model to another region or project and deploying it to a new endpoint, following a dedicated guide in the skill's references folder. Simple questions such as which endpoints exist are routed to a different skill, agent-platform-endpoint-management, and model evaluations are handled by agent-platform-eval-flywheel.
Install the skill (ZIP download or Claude apps)
Download the ZIP for this skill from this page and unzip it. The skill folder is named after the skill and contains SKILL.md at its top level. To make it available across your projects, place the folder in ~/.claude/skills/. To make it available only inside one project, place it in that project's .claude/skills/ folder.
If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the skill there.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills
Set up Google Cloud before your first deployment
The skill assumes the Google Cloud SDK (gcloud CLI) is installed and that you are signed in. Set your project and make sure the Agent Platform and Vertex AI APIs are enabled with billing turned on. The commands use the placeholder variable PROJECT_ID, which you replace with your own project.
The skill also includes a Bash configuration script, scripts/config_gcloud_cli.sh, which the agent can use for gcloud setup. Read it before running it.
- Google Cloud SDK (gcloud CLI), authenticated, with application-default login
- A Google Cloud project with Agent Platform / Vertex AI APIs and billing enabled
- Python 3 (for the cost estimate script scripts/calculate_cost.py)
- A Bash shell (for scripts/config_gcloud_cli.sh)
- A Hugging Face Access Token, only for some Hugging Face models
gcloud auth login
gcloud auth application-default login
gcloud config set project $PROJECT_IDTrigger the skill with natural requests
You do not need special syntax. Describe the task in plain language and mention the model, region or endpoint if you have them. The agent matches your request to the skill's description. Requests that ask only for a list or status of what already exists may go to the endpoint-management skill instead.
If you do not name a model, the agent is required to look up the live catalog and verify the model is deployable before recommending one, rather than suggesting a model from memory.
- "Deploy an open model from Model Garden to a new endpoint in my project."
- "Check the status of my deployment operation."
- "How much would this endpoint cost per hour, and what machine type does it use?"
- "The deploy failed with a quota error. What are my options?"
- "Undeploy the model from endpoint X and clean it up."
- "Copy my tuned model to another region and deploy it to a new endpoint."
How the safety tiers and approvals work
The skill sorts actions into three tiers. Read-only commands such as list, describe and list-deployment-config run without asking. Mutating but reversible actions, deploy and undeploy-model, need your explicit approval. The agent shows a confirmation card with the exact gcloud command, the model identifier, destination project and region, machine type and accelerator, estimated hourly cost, and endpoint display name, then waits for your reply in a later turn.
Deletion is the strictest tier. The agent must explain that deleting an endpoint or model is irreversible and ask you to type "I confirm" or "Yes, delete it". For undeploy-model, the agent first checks that the endpoint and deployed model exist and stops if they do not.
- Tier R (read-only): runs immediately
- Tier M (deploy, undeploy-model): needs approval of a confirmation card
- Tier D (delete): needs typed confirmation
Handle quota and capacity errors
If a deployment fails with QUOTA_EXCEEDED or RESOURCE_EXHAUSTED, the requested hardware is either unavailable in that region or above your project's quota. The error message often suggests another region or machine type. The skill's troubleshooting guide asks the agent to get your confirmation before retrying with the suggested --region or --machine-type, and to recalculate the cost with scripts/calculate_cost.py if the hardware changes.
Safety notes on scripts and network access
This skill runs scripts and uses the network. It calls Google Cloud through gcloud, and some models may need a Hugging Face Access Token. Creating deployments starts running resources that cost money until they are undeployed or the endpoint is deleted. Review the scripts in the skill folder before running them, check the confirmation card carefully, and confirm that the project and region are the ones you intend.
- Check the project ID and region on the confirmation card before approving
- Keep your Hugging Face token out of shared chats and files you do not control
- Undeploy models you no longer need to stop ongoing charges
What people use it for
- Deploying an open-weights model from Model Garden to a dedicated endpoint for testing.
- Estimating the hourly cost of a machine type and accelerator before deploying.
- Diagnosing a quota or resource exhaustion error and retrying in a region the error suggests, after confirming.
- Copying a first-party tuned model to another region or project and deploying it to a new endpoint.
- Undeploying models and deleting unused endpoints to avoid ongoing charges.
Frequently asked questions
How do I install the agent-platform-deploy skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for all projects, or into a project's .claude/skills/ folder for one project. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.
What do I need before using this skill?
You need the Google Cloud SDK (gcloud CLI) signed in with application-default login, a Google Cloud project with Agent Platform and Vertex AI APIs and billing enabled, Python 3 for the cost script, and a Bash shell for the configuration script. Some Hugging Face models also need a Hugging Face Access Token.
Is the agent-platform-deploy skill free?
The skill is licensed under Apache-2.0. Google Cloud resources you deploy are billed to your project, and the skill's own description warns that deployments create costly running resources, so undeploy them when finished.
Can the agent deploy a model without asking me first?
No. Deploy and undeploy actions require you to review a confirmation card and approve it in a later reply. Deleting endpoints or models requires you to type "I confirm" or "Yes, delete it".
What should I do if my deployment fails with a quota error?
Read the error message, which often suggests another region or machine type. Approve the retry only if it works for you, and ask the agent to recalculate the cost if the machine type or accelerator changes.
Which skill should I use to just list my endpoints?
Use agent-platform-endpoint-management for simple questions such as which endpoints exist or which regions have models running. This deploy skill is for actively deploying, checking an operation, costing and cleaning up.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
This skill suits engineers who already run Google Cloud projects with billing and want an agent to handle Model Garden deployments, cost estimates and cleanup with approval gates. If you have no Google Cloud project or do not want to pay for running endpoints, or you only need to list resources, pick the endpoint-management skill or work directly in the console or gcloud.
Download the skill