Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

Analysis · November 7, 2026 · repo

google/skillsai-skillsmcpclaude-codecodex

Google Agent Skills for Claude Code and Codex: Install Two, Not the Whole Catalog

google/skills packs Google Cloud know-how for coding agents. Each skill changes what your agent does on your account, so read it like code before you add it.

Coding agents are good at writing a Cloud Run deploy command and bad at knowing which of three Cloud Run resource types you actually need.

So they guess, from whatever old blog posts the model absorbed, and the guess runs against an account that bills you.

Vendor-written skills are the fix that's spreading, and Google just shipped a big catalog of them. A skill is a folder of instructions (a SKILL.md file plus references) that an agent like Claude Code or Codex loads when a task matches. Instead of guessing, the agent reads the vendor's own playbook.

Google's version, google/skills, sat at #20 on Trendshift's daily board on November 7 with about 23k stars (shields.io, read that day). It covers well over a hundred skills across Google Cloud, Firebase, Google Ads and Analytics, and it installs into several agents with one command.

That one command is the trap. The repo reads like a package, but it behaves like a catalog: each skill is a set of instructions your agent will follow on your account, and several of them reach out to other repos and remote servers. The useful move is to pick the two or three you need and read them first.

What's in the repo

The README calls it a collection of Agent Skills "for Google products and technologies, including Google Cloud." The skills are grouped by area: getting started, AI and ML, infrastructure (a long run of GKE skills), databases and analytics (BigQuery, AlloyDB, Spanner, Cloud SQL), security, web hosting (Cloud Run, Firebase), advertising and analytics APIs, and a set of Well-Architected Framework reviews.

The code is Apache-2.0. The repo has no GitHub releases; it tracks changes through commits (552 on main when we read it), and each skill carries its own version number in its frontmatter (the short settings block at the top of the file).

There are two ways in:

  • Skills only, through the open skills installer: npx skills add google/skills. It asks which skills you want, which agents to configure, and whether to install for one project or globally.
  • Plugins, which the README describes as "Skills + MCP servers" (MCP, the Model Context Protocol, is how agents connect to outside tools and data). Claude Code, Codex and Google's Antigravity CLI (Google's own agent command line) each get their own install path.

The plugin route is where it pays to slow down.

The plugins are mostly pointers

The repo's Claude Code marketplace file, .claude-plugin/marketplace.json, lists 17 plugins under the name google-plugins. Only one of them, google-cloud-developer, lives in this repo.

The other 16 point to separate repositories under the GoogleCloudPlatform organization, each pinned to its own version: bigquery comes from bigquery-data-analytics-plugin at 0.2.5, spanner from spanner-plugin at 0.3.6, looker from looker-plugin at 0.3.11, and so on through AlloyDB, Cloud SQL, Firestore, Dataproc and an Oracle connector.

That's not a flaw. It's how a catalog works. But it means "I installed google/skills" can describe very different setups. Install the BigQuery plugin and you are trusting a different repo, with its own maintainers and its own release history, than the one with 23k stars on it.

The in-repo plugin shows what a plugin adds beyond instructions. Its .mcp.json registers one remote MCP server called developer-knowledge at https://developerknowledge.googleapis.com/mcp, and it sends an API key from an environment variable, DEVELOPERKNOWLEDGE_API_KEY. Its purpose, per the plugin description, is "grounded, up-to-date Google Cloud documentation." In practice your agent can search Google's docs mid-task instead of relying on what the model remembers. The cost is a key to manage and your agent's questions going to a Google endpoint.

Read one skill before you install twenty

The best way to understand what these skills do is to open one. Here's the start of cloud-run-basics, version 1.0.1:

name: cloud-run-basics
description: >-
  Manages Cloud Run services, jobs, and worker pools. Use when you need to deploy applications
  responding to HTTP requests (services), run event-triggered or scheduled tasks (jobs),
  or handle always-on pull-based background processing (worker pools).

The description is the part the agent reads to decide whether the skill applies. It's good: three resource types, each with a plain trigger. That alone fixes the "which Cloud Run thing do I need" problem.

A few lines down sits something most readers will skim past. Under a heading called Attribution, the skill says: "MUST READ: Before executing any gcloud, REST API, or MCP command, you MUST read Usage Attribution." For commands run directly in the terminal, it tells the agent to prefix gcloud with this environment variable:

CLOUDSDK_METRICS_REQUEST_ATTRIBUTION="agent-skills/1.0 (skill:cloud-run-basics)"

and to append the same tag to User-Agent headers. The skill limits this to commands the agent runs itself, "never in generated scripts, Terraform, or user code."

In plain words, the skill asks your agent to label its own Google Cloud calls as coming from this skill. I don't mind that. It's disclosed in the file, it stays out of code you ship, and labeling agent traffic is more honest than hiding it. But it's a good example of why you read a skill before installing it. A skill is not documentation. It's a set of standing orders, and this one changes the commands your agent types.

The same file lists prerequisites the agent will act on, such as gcloud services enable run.googleapis.com cloudbuild.googleapis.com --quiet. That --quiet flag skips confirmation prompts, which is what you want in automation and maybe not what you want the first time an agent touches a new project.

Where to start in a catalog this big

A long list of skills is only useful if you can find the three that match your week. A few groups stand out for builders.

If you're new to a Google Cloud project, the getting-started set covers authenticating and authorizing to Google Cloud (IAM, Google's system for who can do what), onboarding, and a "Foundation Builder" recipe. Read these first if your agent will be the one setting up credentials.

If you deploy apps, Cloud Run Basics and Firebase Basics sit under web and app hosting, and the developer-tools group includes a gcloud CLI skill written for AI agents.

If you own data, the BigQuery group runs from basics to optimization and troubleshooting, alongside AlloyDB, Spanner, Bigtable and Cloud SQL basics.

If you review rather than build, the six Well-Architected Framework skills (cost, operational excellence, performance, reliability, security, sustainability) turn your agent into a checklist reviewer. The security group goes further, with skills such as a Cloud Armor rule upgrader and several Security Command Center and Google SecOps skills.

Pick from one group. You can always come back for another.

Put this into practice

Here's a setup that gets the value without installing the catalog.

1. Install only the skills you'll use, scoped to one project.

npx skills add google/skills

When it prompts, select the two or three skills that match your stack (say, cloud-run-basics and a BigQuery skill), pick the agents you use, and choose project scope rather than global. You need Node.js for npx.

2. Read each SKILL.md before the first run. Look for three things: what the description triggers on, any "MUST" instructions (like the attribution tag), and any command that enables services or skips prompts. Ten minutes per skill is plenty.

3. Add the docs server only if you want it. In Claude Code, the plugin path from the README is:

claude plugin marketplace add google/skills
claude plugin install google-cloud-developer@google-plugins

Then set DEVELOPERKNOWLEDGE_API_KEY in the environment Claude Code runs in. In Codex, the equivalent is codex plugin marketplace add google/skills, then installing from the /plugins browser. In Antigravity CLI, it's agy plugin install https://github.com/google/skills/plugins/cloud/google-cloud-developer.

4. For any other plugin, open its source repo first. The marketplace file names the repo and version for each one. That repo, with its own license and maintainers, is what you're installing, not google/skills.

5. Update on purpose. npx skills update refreshes what you have and offers to remove skills Google has retired. With no releases to pin, a quick look at recent commits before updating is your changelog.

What to expect: the first time you ask your agent to deploy something, it should pick the right Cloud Run resource type, run a prerequisites check, and tag its gcloud calls. If you see the attribution variable in the commands, the skill loaded.

Honest limitations

  • It's a Google Cloud catalog first. Most skills assume a Google Cloud project, the gcloud CLI and the right IAM roles. Without those, they're reading material.
  • Some pieces are preview. The README marks the API Gateway Remote MCP Server skill as Public Preview.
  • No outside pull requests. The contributing guide says Google is "not accepting external pull requests or code contributions," because skills go through internal review. You can file issues, or fork and remix, which the guide encourages.
  • Other Google skills live elsewhere. Android, Flutter, Dart, Firestore, Google Maps and the Agent Development Kit have their own skill repos, linked from the README.
  • No releases. Versions live inside each skill's frontmatter and each plugin's pin, so "which version am I on?" takes a little digging.
  • The docs server needs a key and a network call. Fine for most teams, worth a second thought in locked-down environments.

The real value is the reading

Vendor skills are a good idea. The vendor knows the product's sharp edges, and an agent that knows them too makes fewer expensive mistakes.

But a skill is code in prose form. It tells your agent what to run, what to enable and how to label itself, on an account that bills you. google/skills makes that easy to see, because the files are plain Markdown anyone can open. Install the two you need, read them like you'd read a pull request, and you get Google's playbook without handing your agent the whole library.

Sources: google/skills README; marketplace.json; google-cloud-developer .mcp.json; cloud-run-basics SKILL.md; CONTRIBUTING.md; Trendshift; stars via shields.io. Repo read on main, November 7, 2026.