What it does

agents-cli is a Python CLI and a set of “skills” (structured prompts and context) that plug into coding agents like Gemini CLI, Claude Code, or Codex. It gives them the vocabulary to scaffold ADK projects, run evaluations with LLM-as-judge grading, deploy to Cloud Run or GKE, and wire up observability — all without you memorizing Google Cloud’s service mesh.

You can also use it bare-metal from your terminal: agents-cli scaffold, eval run, deploy, and so on.

The interesting bit

The project treats your coding agent as the primary user interface. Rather than building yet another chatbot, Google built a curriculum for chatbots — teaching them development lifecycle rules, ADK Python patterns, and deployment topology. It’s a manual you don’t have to read because your agent does.

Key highlights

• Scaffold → enhance → upgrade workflow for new or existing ADK projects
• Evaluation suite: trace generation, LLM-as-judge grading, failure-mode clustering, and prompt auto-tuning (in the expanded command set)
• Deployment targets: Agent Runtime, Cloud Run, GKE, plus CI/CD and secrets management
• Standalone or agent-coupled: works via uvx install or npx skills add for IDE integration
• Local development possible with an AI Studio API key; Google Cloud only required for deployment

Caveats

• Pre-GA status with limited support; subject to Google’s Pre-GA offering terms
• Distributed as a pre-built wheel, not source in the repo — you can inspect the zip, but can’t git clone and hack directly
• Requires Python 3.11+, uv, and Node.js just to get started

Verdict

Worth a look if you’re already using Claude Code or Gemini CLI and want to ship ADK agents to Google Cloud without becoming a GCP architect. Skip it if you’re not on the Google stack or prefer to hand-craft your agent infrastructure.