Around 1,800 installs in 24 hours put google-agents-cli-deploy on the skills.sh Trending list, with 533,100 installs since it first appeared on April 21, 2026. It is the deployment skill of agents-cli, Google's CLI for building, evaluating and deploying agents on Google Cloud, and its 1.12.0 release on October 27, 2026 makes ADK Go generally available and adds an experimental extension for agents built with the Claude Agent SDK.
What it does
It kicks in when you ask your agent to "deploy my ADK agent", "set up CI/CD", "configure secrets" or "troubleshoot a deployment", and it covers any framework agents-cli can deploy (ADK, LangChain…). It does not write or evaluate the agent's code: other skills in the same suite handle that.
The useful part is that it gives the agent judgement, not just commands:
- Picking a target. A matrix compares Agent Runtime (Google's managed service for agents), Cloud Run and GKE on scaling, networking, session state, cost model and complexity, and tells the agent to ask you which one you want instead of choosing for you.
- Deploying with permission. The dev flow is: show you the eval scores and test results, ask whether to deploy, and wait for explicit approval before running
agents-cli deploy. The skill says it again in capitals: never deploy without human approval. - Sizing. It explains the defaults (
--cpu 1,--memory 4Gi,--concurrency 8, 0 to 10 instances) and why memory, not CPU, is the first limit: each concurrent request holds its context while it waits on the model, so raising concurrency without memory is the usual cause of OOMs. - Secrets, IAM and rollback. Secret Manager instead of environment variables, which service account needs
secretAccessoron each target, a table of common 403 errors and their causes, and how to roll back: revision traffic on Cloud Run,kubectl rollout undoon GKE, and a redeploy on Agent Runtime, which has no revision-based rollback.
Installing and using it
The skill needs the CLI. Both pieces, per the README and the skill itself:
uv tool install google-agents-cli
npx skills add https://github.com/google/agents-cli --skill google-agents-cli-deployuvx google-agents-cli setup installs the CLI and all seven skills in the suite into your coding agents (Claude Code, Codex, Antigravity CLI…). From there you just ask in plain language. To see what it would do without touching anything, the skill points to agents-cli deploy --dry-run, which in 1.12.0 also validates a Cloud Run deployment against Google Cloud and prints the full gcloud command, and to --no-wait plus --status for Agent Runtime deployments, which take 5 to 10 minutes.
Risks and limits
- Your credentials, your project. The skill declares no
allowed-tools: your agent runsagents-cli,gcloud, Terraform andkubectlas whatever identity is active, and creates resources you pay for. Run it with an account scoped to the dev project. - Broad CI/CD permissions. Setting up the pipeline asks you to grant
roles/secretmanager.adminto the Cloud Build service account. That is what it needs to read the GitHub token, but it is an admin role: review it before applying. - Sharp commands in the troubleshooting table. It lists
terraform force-unlock -forcefor a locked state. If the lock belongs to a pipeline that is still running, forcing it corrupts the state; don't let the agent run it without asking you. - Telemetry. Every agents-cli command writes a log entry to your own project's Cloud Logging and sends Google the version, OS and command name in request headers. Turn it off with
AGENTS_CLI_TELEMETRY=0orDO_NOT_TRACK=1. - A PyPI dependency. The CLI ships as the
google-agents-clipackage; pin the version in CI so changes don't land unreviewed. Version 1.11.0, released on October 20, already brought a breaking change.
The skills.sh page shows the skill passing the Gen Agent Trust Hub, Socket and Snyk audits, and the repository, with 6,278 stars, is licensed Apache 2.0.
Compared with doing it by hand with gcloud
Without the skill, your agent falls back on what it knows about gcloud run deploy, Terraform and the console, and you end up fixing its service accounts and flags. With it, the agent drives a CLI that wraps Terraform, Docker and the deployment into a tested pipeline, and the 403 table saves the usual round of "permission denied". The escape hatch is still there: --dry-run prints the equivalent gcloud command so you can copy it and add what the CLI doesn't expose.
It is worth it if you deploy ADK or LangChain agents on Google Cloud and want your coding agent to follow Google's practices, especially for Agent Runtime, which has no gcloud command. If your agent is just another Cloud Run service behind your own Terraform, it adds less: the valuable bits are the sizing and IAM troubleshooting advice, and those fit in your CLAUDE.md.
Sources
- google-agents-cli-deploy — skills.sh
- agents-cli: the CLI and skills for building agents on Gemini platform — Google (GitHub)
- google-agents-cli-deploy SKILL.md (v1.12.0) — Google (GitHub)
- agents-cli release notes (1.12.0) — Google (GitHub)
- Trending skills (24 h) — skills.sh