google/agents-cli

The CLI and skills that turn any coding assistant into an expert at creating, evaluating, and deploying AI agents on Google Cloud.

What is agents‑cli?

agents‑cli is a command‑line tool (and a set of “skills” you can add to AI‑powered coding assistants) that helps you build, test, and deploy AI agents that run on Google’s Gemini platform. It wraps the lower‑level Agent Development Kit (ADK) and Google Cloud services so you don’t have to manually stitch together many separate CLIs.

Who is it for?

Developers who want to create “agent” applications (LLM‑backed bots that can call tools, maintain state, etc.) and run them on Google Cloud.
Coding assistants such as Antigravity CLI, Claude Code, or OpenAI Codex – the tool can be installed as a skill that lets those assistants issue the right commands for you.
Teams that need a repeatable, CI/CD‑ready workflow for agent projects (scaffolding, evaluation, deployment, observability).

Core features (as described in the README)

Area	What you get
Scaffolding	`agents-cli create <name>` creates a new project with the ADK‑compatible layout. `scaffold enhance` adds deployment or CI/CD to an existing repo, and `upgrade` bumps the project to the latest CLI version.
Development helpers	`run "prompt"` lets you invoke the agent locally, `install` resolves dependencies, and `lint` runs Ruff for code quality.
Evaluation suite	Commands under `agents-cli eval …` let you generate inference traces, grade them with built‑in metrics, synthesize multi‑turn scenarios, compare runs, cluster failure modes, and even auto‑tune prompts (`optimize`).
Deployment & publishing	`deploy` pushes the agent to Google Cloud (Cloud Run, GKE, or the Gemini‑managed runtime). `publish gemini` registers the agent with the Gemini platform. Infrastructure helpers (`infra single-project`, `infra cicd`) set up the required Google Cloud resources.
Observability	Integration with Cloud Trace, Cloud Logging, and third‑party monitoring is baked in via the `observability` skill.
Authentication	`agents-cli login` authenticates you with Google Cloud or AI Studio; `login --status` shows the current token.
Telemetry control	By default a minimal usage log is sent to your own Google Cloud project; you can disable it with `AGENTS_CLI_TELEMETRY=0` or `DO_NOT_TRACK=1`.

How it works under the hood

ADK Python API – the CLI generates code that uses the Agent Development Kit, which defines agents, tools, callbacks, and state handling.
Gemini platform – when you deploy or publish, the CLI provisions the Agent Runtime on Gemini, or standard Cloud Run / GKE services, handling IAM, secrets, and CI/CD pipelines.
Evaluation framework – a built‑in dataset format and “LLM‑as‑judge” grading system let you measure quality (e.g., correctness, relevance) and feed the results back into prompt optimization.
Skill layer – the same commands are exposed as skills that can be installed into external coding agents (via npx skills add google/agents-cli). Those agents can then issue the CLI commands on your behalf.

Getting started (quick recap)

Prerequisites – Python 3.11+, uv (a fast Python package manager), and Node.js.
Install – uvx google-agents-cli setup (or add the skill to a coding agent with npx skills add google/agents-cli).
Create a project – agents-cli create my‑agent.
Iterate – run prompts, lint, and evaluate locally.
Deploy – agents-cli deploy to push to Google Cloud, then agents-cli publish gemini to register.

Why it matters

End‑to‑end workflow – eliminates the need to manually combine ADK, Cloud Run, GKE, and logging setups.
Agent‑centric evaluation – provides a ready‑made framework for multi‑turn testing and prompt tuning, which is often missing from generic LLM toolkits.
Integration with coding assistants – the skill model lets AI‑powered IDE helpers become the “operator” for your agent projects, streamlining the developer experience.

Where to learn more

Full documentation site: https://google.github.io/agents-cli/
Quick‑start tutorial: https://google.github.io/agents-cli/guide/quickstart-tutorial/
Release notes, issue tracker, and PyPI package are linked in the README.

All information above is taken directly from the repository’s README; no additional features have been assumed.