agent-platform-deploy Skill for Claude: Deploy Model Garden Models With gcloud

Makes Claude deploy open models from Google's Model Garden to endpoints with gcloud, with a cost estimate and your approval first.

Installs: 9.3K; on skills.sh
Stars: 21.8K; on GitHub
Files: 7; besides SKILL.md
Updated: Nov 2026; last repo commit

This skill is published by Google in its own google/skills repo. It makes Claude use the gcloud ai commands to browse the Model Garden catalog, check which machine types a model supports, estimate the hourly cost, deploy to an endpoint, check progress, and later undeploy and delete what it created.

Deploying large models uses paid GPU compute. SKILL.md sets three tiers: reads need no confirmation, deploys and undeploys need your explicit confirmation after a card showing the exact command, region, hardware and hourly cost, and deletes need you to type a confirmation. You need gcloud installed and signed in to a project.

Good for

Google Cloud users who want to deploy an open model such as Gemma or Llama to an endpoint
Teams that need a cost estimate and a clean teardown so GPU endpoints do not keep billing
Claude Code users with gcloud installed and signed in

Install agent-platform-deploy

Claude Code

Copy this prompt.

Install the agent-platform-deploy skill in Claude Code. Run this command:

npx skills add google/skills --skill agent-platform-deploy

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

Pick one route. The README installs the whole google/skills repo with npx skills add google/skills; the first command names only this skill. The repo's Google Cloud plugin (google-cloud-developer) does not contain this skill, so there is no plugin route. SKILL.md has you run gcloud auth login, gcloud auth application-default login and gcloud config set project $PROJECT_ID first. The cost script needs python3. There is no Claude app upload route because the skill needs gcloud on your machine.

Built from the package data we hold. Not yet tested by us. Source.

What's in the skill

When it loads

Deploy open models or custom weights from Model Garden to Agent Platform endpoints, check the status of an in-progress deployment operation, or clean up resources by undeploying models and deleting endpoints. Use when asked to actively deploy a model, list the Model Garden CATALOG of available models, check if a specific model is deployable (`gcloud ai model-garden models list-deployment-config`), query deployment cost, troubleshoot deployment errors (like quota limits), or undeploy/clean up endpoints. Also use when copying and deploying a 1P Tuned Model. Don't use for pure listing/discovery questions of the form "is X deployed?", "list my endpoints", or "which regions have models running?" — for those use `agent-platform-endpoint-management`. Don't use for running model evaluations (use `agent-platform-eval-flywheel` skill).

Sections in SKILL.md

1P Tuned Model Copy & Deployment
Safety & Confirmation Tiers (CRITICAL)
1. Prerequisites
2. Discovering Deployable Models
2.1 Region Availability Check (Gemini + LoRA only)
3. Deploying a Model
Example: Deploying an open-weights model from Model Garden
1P Tuned Model Cross-Region Copy and Deployment
4. Checking Deployment Status
Verifying Deployment
5. Undeploying and Cleaning Up
6. Troubleshooting

Key facts

Author: Google (vendor's own repo google/skills)
Repo: google/skills
Installs on skills.sh: 9,284 (Nov 2, 2026)
GitHub stars: 21,764 (Nov 2, 2026, whole repo)
License: Apache-2.0 (crawl data)
Last repo commit: Nov 1, 2026
What installs: 8 files including SKILL.md: 5 reference notes, calculate_cost.py and config_gcloud_cli.sh
Skill version: 1.0.4
Needs: gcloud, signed in with gcloud auth login and a project set; python3 for the cost script
skills.sh audits: Gen Agent Trust Hub pass, Socket pass, Snyk pass (Nov 2, 2026)

What it includes

Reads run freely; deploys need a confirmation card; deletes need typed confirmation
Tier R is list, describe and list-deployment-config. Tier M is deploy and undeploy-model. Tier D is delete. Claude must end its turn after asking and wait for your reply.
Safety tiers
gcloud ai model-garden models list and list-deployment-config for hardware options
Claude must not name a model version from memory. It queries the live catalog, verifies the exact ID is deployable, and quotes it verbatim.
Runs code · Discover models
Estimates the hourly list price for a machine type before any deploy
Takes --machine-type, such as g2-standard-48. If the type is unknown it exits non-zero and Claude falls back to the Google pricing page.
Runs code · scripts/calculate_cost.py
gcloud ai model-garden models deploy ... --asynchronous after you approve the card
Optional flags include a Hugging Face token for gated models and a reservation affinity. It checks region availability first for Gemini and LoRA deploys.
Runs code · Deploy a model
gcloud ai operations describe, run once, then a test prediction when serving
No sleep or polling loops. The test prediction is a curl call to the endpoint's dedicated DNS with a gcloud access token.
Runs code · Check status
Undeploys the model, then deletes the endpoint and model after typed confirmation
Uses gcloud ai endpoints undeploy-model, endpoints delete and models delete. SKILL.md warns that a deployed model keeps charging until undeployed.
Delete · Undeploy and clean up
references/copy_deploy_guide.md: copy a 1P tuned model across regions or projects
Uses scripts/config_gcloud_cli.sh to create and activate a gcloud configuration, and can add IAM bindings. It names autopush, staging and prod environments.
Runs code · Tuned model copy and deploy
references/troubleshooting.md: quota and resource-exhausted errors, hardware fallback
SKILL.md also tells Claude to offer another region or smaller hardware, with a fresh cost estimate, when a deploy fails on quota.
Troubleshooting

From the official docs

Permissions

High risk

It can start paid GPU endpoints and delete endpoints and models in your Google Cloud project. Use a project and role limited to this work, and read each confirmation card.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

It can create paid GPU endpoints
A deployed model keeps charging for its allocated compute until it is undeployed, even with no requests (SKILL.md). The charge lands on the Google Cloud project you deploy into (our reading). SKILL.md requires an hourly list-price estimate and your confirmation first, tells Claude to end its turn after the confirmation card, and warns that your real bill may differ from the list price. The estimate comes from scripts/calculate_cost.py, or Google's pricing page if the machine type is unknown. github.com
It can delete endpoints and models
Deletes are irreversible. SKILL.md asks you to type a confirmation such as "I confirm" before running them. github.com
The copy guide can change IAM and gcloud config
The tuned-model guide runs gcloud projects add-iam-policy-binding for a service agent role, can grant roles/aiplatform.admin to your user on the destination project, and creates and activates a gcloud configuration through config_gcloud_cli.sh. Do not approve these unless you meant to. github.com
Gated models and test prompts use tokens
For gated Hugging Face models the deploy command takes --hugging-face-access-token. The status check sends a prompt to the endpoint's dedicated DNS with a gcloud access token in an Authorization header. github.com

Technical details

Talks to	Google Cloud APIs and your endpoint's dedicated DNS name, through gcloud and curl. cloud.google.com (pricing page fallback)
Sign-in and keys	gcloud sign-in (gcloud auth login and application-default login); the skill uses gcloud auth print-access-token for the test request.
Scopes	Runs gcloud ai commands that list models, deploy models to paid endpoints, undeploy them and delete endpoints and models Runs python3 scripts/calculate_cost.py for the hourly estimate Sends a test prompt to your endpoint with curl and gcloud auth print-access-token Tuned-model copy flow: can add IAM bindings and create a gcloud configuration

Questions

What does agent-platform-deploy do?

It makes Claude deploy open models or custom weights from Google's Model Garden to Agent Platform endpoints with gcloud. It also checks deployment status and undeploys and deletes what you no longer need.

How do I install it in Claude Code?

Use the install box at the top. The README installs the whole google/skills repo, and the box also shows a command for this skill only. You also need gcloud installed and signed in to a project, because the skill drives gcloud and curl.

Will it spend money without asking?

SKILL.md says no. Claude must show a card with the exact command, region, hardware and estimated hourly cost, and wait for your reply before any deploy. The estimate is a list price, so your real bill may differ.

How do I stop the charges?

Undeploy the model and delete the endpoint. SKILL.md warns that failing to undeploy keeps charging for the compute even with no requests. Deletes need typed confirmation.

Is it made by Google?

Yes. It sits in Google's own google/skills repo.

What about checking whether a model is already deployed?

Its description sends pure listing questions, such as "is X deployed?" or "list my endpoints", to a different skill, agent-platform-endpoint-management. Model evaluations go to agent-platform-eval-flywheel.

Does it work in the Claude app?

The README does not document the Claude app. The skill runs gcloud and curl on your machine, so use a client that can run commands.

Is it safe?

It can start paid GPU endpoints and delete resources, and it runs gcloud, curl and two bundled scripts (calculate_cost.py and config_gcloud_cli.sh). Each deploy needs your approval of a card with the exact command and hourly cost, and each delete needs a typed confirmation. The configuration script is called from the tuned-model copy guide. Read each card before you approve.

Information

updated Nov 2, 2026

Repository: google/skills
Folder: skills/cloud/agent-platform-deploy

google/skills skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 5

SKILL.md · github.com
google/skills README · github.com
copy_deploy_guide.md · github.com
undeploy_guide.md · github.com
config_gcloud_cli.sh · github.com