Google’s `gcloud` skill: a safety gate before your agent touches GCP (2026)

The short verdict

Install Google’s gcloud skill if your coding agent regularly proposes Google Cloud CLI commands and you want a reviewable safety step before anything touches a project. It is a small, focused member of Google’s official google/skills repository: its job is to make the agent validate the exact leaf command, reduce command output, preserve project scope, and stop for authorization before risky operations. 1 2

Skip it if your work is mainly Google Cloud client-library code, raw REST or gRPC calls, or a non-GCP stack. The skill explicitly does not cover those paths. 3

One honesty check before the numbers: this is a strong officialness and adoption pick, not a claim that google/skills is in GitHub’s current weekly Trending list. The public Trending page did not show the repository in the snapshot I checked, and I could not verify a reliable seven-day star delta. Treat the totals below as reach signals, not proof of quality or weekly breakout.

Why this pick made the cut

google/skills is Google’s public repository for Agent Skills covering Google products and technologies, including Google Cloud. The repository says it is under active development; GitHub’s API reports an Apache 2.0 license, about 17.1K stars, and a last push on August 7, 2026. 1 4

The catalog is broad: the skills.sh page currently lists 113 skills and about 296.1K total installs for the repository. The narrower gcloud entry shows about 4.9K installs, a first-seen date of June 1, 2026, and automated “Pass” labels from Gen Agent Trust Hub, Socket, and Snyk. Those labels are useful supply-chain signals; they are not a human correctness review or a guarantee that a command is safe for your project. 2 5

The reason to start with gcloud rather than install the whole catalog is its narrow payoff: it changes how an agent handles a high-consequence interface. It does not replace the Google Cloud CLI, grant permissions, or inspect your infrastructure for you.

What the `gcloud` skill changes

1. It makes leaf-level help the first step

The skill treats all remembered gcloud syntax as stale. Before proposing or executing a command, the agent must ask the local CLI for help on the exact leaf command—for example, gcloud help compute instances list. Help for the parent group is not enough, and web search is not an allowed fallback for command syntax. 3

That is a practical anti-hallucination rule. The agent can still misunderstand your goal, but it has less room to invent a flag that your installed SDK does not support.

2. It treats output size as an operational risk

The skill forbids unconstrained list commands. A list must carry at least one reduction control such as --limit, --filter, or --format; if the schema is unclear, it recommends inspecting one item first with a narrow JSON response. 3

The schematic shape is:

gcloud <GROUP> <RESOURCE> list --limit=1 --format=json

That is a pattern, not a command to paste unchanged. The agent should validate the real leaf command first, then choose the fields and filter that match the resource.

3. It keeps scope and authorization visible

The guidance requires an explicit --project=<PROJECT_ID> for resource queries and mutations rather than silently relying on an active default. It also calls for explicit location flags where a command needs a region or zone, single-command execution without shell chaining, and --quiet for non-interactive runs. 3

Deletes, updates, removals, IAM changes, billing operations, organization changes, KMS operations, and API enabling are treated as human-authorized territory. If a command supports --dry-run or --validate-only, the skill says to use that preview before the real operation. 3

Install it and run a safe first test

Prerequisites

Node.js with npm/npx, or an agent host that can install the Agent Skills format.
The Google Cloud CLI if you want the agent to execute commands rather than only plan them. If it is not installed, the skill points to Google’s official CLI installation guide. 3
A Google Cloud identity and a deliberately chosen project. For local interactive work, the skill documents gcloud auth login; headless environments can use gcloud auth login --no-browser. For application code, it distinguishes Application Default Credentials from the CLI user login. 3
Enough project permissions for the read-only task you will test. Installing the skill does not provide IAM access.

Portable install

Run the repository’s documented install command:

npx skills add google/skills

When the installer presents the catalog, choose gcloud. Google’s README also documents plugin routes for Claude Code, Codex, and Antigravity CLI; these routes install Google’s plugin bundle rather than proving that every host exposes the gcloud leaf in exactly the same way. Cursor and Cline support were not independently confirmed from this repository’s README, so check your host’s skill discovery after installation. 1

After installation, ask the host to list or explain its loaded skills. If it has a reload command, reload before testing. The repository is active enough that the installed copy, the catalog’s current default, and the copy actually loaded by a running agent should not be assumed to be identical.

First test: a read-only, bounded inventory

Use a disposable or low-risk project and give the agent a prompt like this:

Use the gcloud skill. I need a read-only inventory of running Compute Engine
instances in project <PROJECT_ID>.

1. First validate the exact leaf syntax with:
   gcloud help compute instances list
2. Preserve --project=<PROJECT_ID> in the proposed command.
3. Use a bounded result: a limit, a status filter, and a narrow JSON projection
   containing only name, zone, and status if the validated syntax supports them.
4. Check whether this leaf command supports --dry-run or --validate-only.
5. Do not execute anything yet. Show the validated command and explain what it
   will return.

A good response should show the help lookup first, keep the project explicit, and avoid dumping full instance metadata. If it jumps straight to a remembered command or returns an unbounded list, the test has already found a reason not to trust the setup.

Second test: make a risky request stop cleanly

Then test the boundary rather than a happy path:

Use the gcloud skill. I want to remove an unused resource in project
<PROJECT_ID>.

Do not delete or update anything. First identify the exact leaf command, run
leaf-level help, verify the project and resource, check for --dry-run or
--validate-only, and give me a read-only inspection command. Tell me exactly
which step would require my explicit authorization.

The expected result is not a clever deletion command. It is a pause: validate the leaf syntax, inspect the target narrowly, and ask for authorization before a destructive operation. That is the behavior this skill is designed to add.

What the community signal actually says

There is no mature star-rating or review corpus for this individual skill that I could verify. The most useful public feedback is about coverage gaps, not applause.

In open issue #105, unrealandychan argues that agents fall back to gcloud or adjacent skills for Cloud Storage while missing storage-specific decisions such as uniform bucket-level access, signed URLs, storage classes, lifecycle rules, and CORS. That is a concrete request for a dedicated GCS skill—not evidence that the existing gcloud skill is broken. 6
In open issue #106, the same contributor describes a Day-2 operations gap around Cloud Monitoring, Logging, SLOs, tracing, and alerting. Again, this is user demand for narrower coverage, and a useful warning not to mistake a safe CLI wrapper for a complete operations runbook. 7
A Google Cloud Community walkthrough praises the repository’s narrow-scope design and documents both global and single-skill installation patterns. That is an attributable practitioner walkthrough, not an independent benchmark or review panel. 8

One adjacent issue deserves attention if you use Google’s Agent Platform registry skills. In #183, a contributor points out that a successful upload, a new registry revision, the registry’s default revision, and the revision actually loaded by a long-lived agent are four different facts. The requested reporting pattern ends with runtime_loaded_revision: unknown unless the loader provides evidence. This issue concerns the registry skill rather than the local gcloud leaf, but the lesson generalizes: an installed or published skill is not proof that your current agent session loaded the version you intended. 9

When to use it—and when to skip it

Use gcloud when:

Your agent writes, reviews, or executes Google Cloud CLI commands.
You want every command proposal tied to the locally installed CLI’s leaf-level help.
You need bounded discovery instead of full resource dumps.
You want explicit project scope and a human checkpoint around destructive, IAM, financial, or governance changes.

Sandbox or skip it when:

You are writing client-library code or raw REST/gRPC requests; the skill says those are out of scope. 3
You need Cloud Storage, observability, or another service’s deep domain guidance. The open requests above show why a general CLI safety layer is not a substitute for a service-specific skill. 6 7
Your host cannot show which skill is loaded or cannot reload after installation. Verify the loaded copy before trusting it with consequential work.
You cannot provide a real project, identity, and permission boundary for the first test. Do not “test” against a production default just because the CLI is already authenticated.

Quick reference

Pick	`gcloud`, from Google’s official `google/skills` catalog
Best for	Agents that propose or execute Google Cloud CLI commands
Officially documented routes	`npx skills add google/skills`; Google also documents Claude Code, Codex, and Antigravity plugin routes. 1
Adoption signal	About 4.9K installs for `gcloud`; about 296.1K installs across the repository on skills.sh. 2 5
Repository signal	About 17.1K GitHub stars; Apache 2.0; last push observed August 7, 2026. 1 4
Main mechanism	Leaf-level `gcloud help`, narrow output, explicit scope, and authorization gates
Main limitation	It is not client-library/API guidance, a permission grant, a runtime-version attestation, or a substitute for service-specific skills
First move	Install only `gcloud`, test a read-only inventory in a named project, and inspect the exact loaded skill before widening access

The sensible rollout is small: install gcloud, run the read-only inventory prompt, and judge the response by whether it validates the local leaf command and keeps the result bounded. If you want an agent to manage production GCP, this is a useful safety gate—but it is still a gate, not a second operator.