Google Agent Platform Inference Skill for Claude Code: What It Does and How to Install It

Makes Claude run test prompts and write inference code for Gemini and open models (Llama, DeepSeek, Qwen) on Google Cloud Agent Platform.

Installs: 8.6K; on skills.sh
Stars: 21.8K; on GitHub
Files: 8; besides SKILL.md
Updated: Nov 2026; last repo commit

Ask Claude to run a prompt on a Gemini or open model and this skill has it collect the project ID, region, model ID and SDK. Claude then shows a Yes/No confirmation card and waits for your reply before it runs anything. It also covers custom endpoints (tuned Gemini, self-deployed open models, legacy models) and a troubleshooting section for 429, 400 and 404 errors.

It needs a Google Cloud project, the gcloud CLI, Python 3 and the SDKs google-genai, openai and google-cloud-aiplatform. Calls go to your project, so they can cost money or use quota. SKILL.md says this is why it asks first. It is not for deploying models or running evaluations.

Good for

Developers calling Gemini or open models from their own Google Cloud project
People who want working Python for the Google GenAI SDK or the OpenAI SDK
Anyone debugging 429, 400 or 404 errors from Agent Platform models

Install agent-platform-inference

Claude Code

Copy this prompt.

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

npx skills add https://github.com/google/skills --skill agent-platform-inference

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, gcloud CLI and a Google Cloud project, Python 3.

Pick one route. The README command npx skills add google/skills lets you pick skills when prompted; the command here installs only agent-platform-inference. No plugin in the repo marketplace holds this skill. The skill needs gcloud, Python 3 and a Google Cloud project, so there is no Claude app upload route.

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

What's in the skill

When it loads

Connects to and performs inference with Google Cloud Agent Platform GenAI models, including First-Party Gemini models and Third-Party OpenMaaS models (Llama, DeepSeek, Qwen, etc.). Use when asked to perform inference, ask a model a question, run a test prompt, execute chat completions, or generate code for calling Gemini or OpenMaaS models, authenticate with GenAI SDK, OpenAI SDK, or legacy Agent Platform SDK, configure base URLs and global/regional endpoints, or troubleshoot 429 Resource Exhausted (DSQ), 400 User Validation, or 404 Not Found errors. Don't use for deploying models to endpoints or for running model evaluations.

Sections in SKILL.md

Safety & Confirmation Tiers (CRITICAL)
Phase 0: Environment Setup
Parameter Grounding & Clarification Protocol (CRITICAL)
Workflow Decision Tree
0.5 Region Availability Check for Publisher Endpoints (Gemini + LoRA base)
1. Gemini Models
Choosing the Right SDK
Installation
Python Example (GenAI SDK - Preferred)
Alternative: OpenAI SDK (Chat Completions)
Legacy: Agent Platform SDK
2. OpenMaaS Models (Llama, DeepSeek, Qwen, etc.)
Installation
Authentication for OpenAI SDK
Configuration (Base URL)
Python Example (OpenMaaS - Chat Completions)

Key facts

Author: Google (the google/skills repo; marketplace owner is Google LLC)
Repo: google/skills
License: Apache-2.0 (LICENSE and the repo metadata agree)
Installs on skills.sh: 8,554 (Nov 2, 2026)
GitHub stars: 21,790 for the whole repo (Nov 2, 2026)
Last repo commit: Nov 1, 2026
Audits on skills.sh: Gen Agent Trust Hub: pass, Socket: pass, Snyk: pass
What it includes: 9 files: SKILL.md, 6 Python sample scripts, requirements.txt and verify_all.sh
Needs: A Google Cloud project, gcloud, Python 3, and the google-genai, openai and google-cloud-aiplatform packages

What it includes

Shows project, region, model, SDK and prompt as a Yes/No card and waits for your reply.
SKILL.md says inference must not run in the same turn as the card, and the answer must repeat the model, SDK, project and region used.
Read · Confirmation card
Asks which model, SDK, project ID and region to use instead of guessing.
SKILL.md says never to default to a model.
Ask for missing details
Runs gcloud auth login, application-default login and enables the aiplatform API.
Also pip installs the SDKs from scripts/requirements.txt when the imports fail.
Runs code · Environment setup
Runs every sample script in scripts/ and reports pass or fail.
Each of the 6 scripts makes a real model call, using your Google credentials.
Runs code · scripts/verify_all.sh
Sends one minimal generateContent request to see if the model exists in your region.
Uses curl with a gcloud access token. A 404 means the model is not offered there.
Runs code · Region check for Gemini
Python with google-genai (preferred), the OpenAI SDK or the legacy vertexai SDK.
Files: gemini_genai_sdk.py, gemini_openai_sdk.py, gemini_vertexai_sdk.py.
Try: “Ask gemini-2.5-pro why the sky is blue”
Runs code · Gemini examples
Python for Llama, DeepSeek, Qwen and similar models, mostly via the OpenAI SDK.
Files: openmaas_openai_sdk.py, openmaas_genai_sdk.py, openmaas_vertexai_sdk.py.
Runs code · OpenMaaS examples
Calls a model on your own numeric endpoint ID with the method that matches it.
Tuned Gemini uses generateContent, self-deployed open models use chat/completions, legacy models use predict. Uses gcloud ai endpoints list and describe.
Runs code · Custom endpoints
Explains 429 Resource Exhausted, 400 User Validation, 404 Not Found and empty replies.
Covers backoff for shared quota, Provisioned Throughput, and enabling Llama models in Model Garden.
Read · Troubleshooting

From the official docs

Permissions

Medium risk

What it can do

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

Watch out for

Model calls can cost money
SKILL.md requires a Yes/No confirmation before inference to prevent unexpected cost or quota use. Read the card before you say Yes, and set a budget alert on the project. github.com
verify_all.sh runs six real model calls
The script runs every Python file in scripts/, and each one calls a model with your default Google credentials. If a script fails it runs that script a second time to show the output. It also runs pip install -q -r scripts/requirements.txt if the SDK imports fail. Skip it unless you want that check. github.com
It changes your Google Cloud setup
The setup steps sign in with gcloud and enable the aiplatform API on the active project. Check which project gcloud points at first. github.com
Model IDs in the skill may be out of date
SKILL.md says its model IDs are examples that may be incomplete or outdated and tells Claude to check the Google docs before writing code. github.com

Technical details

Turning on read-only	SKILL.md calls the skill read-only, but inference still spends quota. Its confirmation step is the control.
Talks to	aiplatform.googleapis.com and regional aiplatform.googleapis.com hosts (model calls). Dedicated endpoint hosts on prediction.vertexai.goog (custom endpoints). PyPI (pip install of the SDKs)
Sign-in and keys	Uses your gcloud sign-in and Application Default Credentials. The sample scripts read the default project from those credentials. SKILL.md names no API key.
Scopes	Signs in to Google Cloud with gcloud auth login and application-default login Enables the aiplatform.googleapis.com API on your project Runs Python scripts and curl calls that send prompts to Google Cloud models Installs Python packages with pip when imports are missing

Questions

What does the agent-platform-inference skill do?

It makes Claude run a prompt on a Gemini or open model (Llama, DeepSeek, Qwen) on Google Cloud Agent Platform, or write Python for it. It also covers custom endpoints and 429, 400 and 404 errors.

How do I install it?

Use the install box at the top of this page. The README documents the skills CLI and says you can pick specific skills when prompted. No plugin in the repo marketplace holds this skill.

Does it need a Google Cloud account?

Yes. SKILL.md has Claude run gcloud auth login and gcloud auth application-default login, and enable the aiplatform API on your project. Model calls then use that project.

Will it spend money without asking?

SKILL.md tells Claude to show a Yes/No card with project, region, model, SDK and prompt, and to wait for your reply before running inference. It says the card exists to prevent unexpected cost or quota consumption. Review the card each time and set a budget alert on the project.

Is it safe?

It signs in to your Google Cloud account, enables an API, installs Python packages and sends prompts to a model. Use a test project and a budget alert.

Does it work in the Claude app?

The install box has no Claude app upload route. The skill runs gcloud, Python and curl on your machine.

Can it deploy a model or run evaluations?

No. The description says not to use it for deploying models to endpoints or running model evaluations. Deployment belongs to the agent-platform-deploy skill in the same repo.

Is it an official Google skill?

It lives in Google's own google/skills repo, whose README says it holds Agent Skills for Google products.

Information

updated Nov 2, 2026

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

google/skills skills.sh

Sources

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

Show all 4

google/skills README · github.com
agent-platform-inference SKILL.md · github.com
google/skills marketplace.json · github.com
verify_all.sh · github.com