# Agents CLI

> Scaffold, evaluate, deploy and publish ADK agents from one CLI — and hand the same commands to your coding assistant as skills

- Category: Agents & Automation / Agent Frameworks & Builders
- Language: Python
- Licence: Apache-2.0
- Repository: https://github.com/google/agents-cli
- Website: https://google.github.io/agents-cli/
- Canonical page: https://ai-tldr.dev/tools/agents-cli/

## In short

Agents CLI (`agents-cli`) is Google's command-line tool for the lifecycle of an agent built on the Agent Development Kit: scaffold a project, run it, evaluate it, deploy it to Google Cloud, register it with Gemini Enterprise, and observe it in production. It sits in the Agent Frameworks & Builders category of the Agents & Automation tools and is open-source, written in Python and licensed Apache-2.0.

## Overview

Agents CLI (`agents-cli`) is Google's command-line tool for the lifecycle of an agent built on the Agent Development Kit: scaffold a project, run it, evaluate it, deploy it to Google Cloud, register it with Gemini Enterprise, and observe it in production. It is explicitly not a coding agent and not a replacement for ADK — ADK is the framework you write the agent in, and `agents-cli` is the tooling around it.

Its second half is the part that distinguishes it. `agents-cli setup` installs a set of agent skills into whatever coding assistant you already use — Antigravity CLI, Claude Code, Codex, or another — so that assistant knows the ADK Python API, the scaffolding commands, the evaluation methodology, the deployment targets and the observability wiring without you having to teach it. Seven skills cover workflow, ADK code, scaffolding, eval, deploy, publish and observability. The CLI also works standalone: every command runs directly from a terminal, and the skills are a convenience layer, not a requirement.

Evaluation is unusually well covered for a scaffolding tool. Beyond `eval run`, the CLI can synthesise multi-turn eval scenarios for a local agent, grade generated traces against named metrics, compare two result files, cluster failure modes out of grade results, and auto-tune agent prompts from the eval data. Local development — `create`, `run`, `eval` — works against an AI Studio API key without a Google Cloud project; Cloud is needed for deployment and the platform features. The project is Apache-2.0 licensed and published to PyPI as `google-agents-cli`. It is the successor to Agent Starter Pack, which moved to maintenance mode and points new projects here.

## What it does

- Project scaffolding with `create`, plus `scaffold enhance` to add deployment and CI/CD to an agent project that already exists
- Evaluation suite: run and grade traces, synthesise multi-turn eval scenarios, compare two result files, cluster failure modes, and auto-tune prompts from eval data
- Deployment to Agent Runtime, Cloud Run or GKE, with `infra single-project` and `infra cicd` provisioning the surrounding infrastructure and pipelines
- `publish gemini-enterprise` registers a finished agent with Gemini Enterprise
- Seven agent skills installed into Claude Code, Codex, Antigravity CLI or any other coding assistant, covering the ADK API, scaffolding, eval, deploy, publish and observability
- Observability wiring for Cloud Trace and logging, plus third-party integrations

## Getting started

You need Python 3.11+, uv and Node.js. One command installs both the CLI and the skills; local development needs only an AI Studio API key.

1. Install the CLI and skills

   `setup` installs the CLI and pushes the skills into the coding assistants it finds.

```bash
uvx google-agents-cli setup
```

2. Or install just the skills

   If you would rather let your coding agent handle the rest, add the skills on their own.

```bash
npx skills add google/agents-cli
```

3. Create and run an agent

   Scaffold a project, install its dependencies, then drive it with a single prompt. Authentication goes through Google Cloud or AI Studio.

```bash
agents-cli login
agents-cli create my-agent
cd my-agent
agents-cli install
agents-cli run "summarise the attached report in three bullets"
```

4. Evaluate before you ship

   Synthesise eval scenarios, run and grade them, then look at where it went wrong — `eval analyze` clusters the failure modes rather than leaving you to read every trace.

```bash
agents-cli eval dataset synthesize
agents-cli eval run
agents-cli eval analyze
agents-cli eval compare before.json after.json
agents-cli eval optimize
```

5. Deploy and publish

   Provision infrastructure, deploy to Google Cloud, and register the agent with Gemini Enterprise.

```bash
agents-cli infra single-project
agents-cli deploy
agents-cli publish gemini-enterprise
```

## When to use it

- Take an ADK agent from an empty directory to a deployed, registered Gemini Enterprise agent without assembling the CI/CD and infrastructure yourself
- Give Claude Code or Codex working knowledge of ADK and Google Cloud deployment so it can build agents for you instead of guessing at the API
- Put a real evaluation loop around an agent — synthesised scenarios, graded traces, clustered failure modes, prompt auto-tuning — rather than eyeballing outputs
- Add production deployment, observability and CI/CD to an agent project that was prototyped without them