google/agents-cli: Skills and Commands That Teach Your Coding Agent to Build ADK Agents

The gap agents-cli fills between a coding agent and Google Cloud

A general coding assistant knows how to write Python. It does not know which ADK callbacks exist, which eval metrics the platform supports, or which infrastructure a single-project deployment needs. The README frames the project as exactly that gap: it "gives your coding agent the skills and commands to build, scale, govern, and optimize enterprise-grade agents." The audience is a developer who writes agent code in an editor but does not want to memorize the surface area of the Gemini Enterprise Agent Platform. The README also answers the most likely misreading directly: agents-cli "is a tool for coding agents, not a coding agent itself." If you are shopping for an autonomous coding assistant, this is the wrong repository. If you already have one and want it to stop guessing about ADK, this is the layer that supplies the missing knowledge.

Skills shipped into the assistant, commands run by you

The mechanism is a two-part split. The skills are prose and reference material that the coding agent loads, and the CLI is the executable half. The repository root reflects that split: a `skills/` directory, a `src/` directory, plus `gemini-extension.json` and `plugin.json` manifests, which is how the same package can be surfaced to more than one assistant. Seven skills are listed, each mapped to a phase: `google-agents-cli-workflow` for lifecycle and model selection, `google-agents-cli-adk-code` for the ADK Python API, `google-agents-cli-scaffold` for `create`, `enhance` and `upgrade`, `google-agents-cli-eval` for metrics and LLM-as-judge, `google-agents-cli-deploy` for Agent Runtime, Cloud Run, GKE, CI/CD and secrets, `google-agents-cli-publish` for Gemini Enterprise registration, and `google-agents-cli-observability` for Cloud Trace and logging. The command surface mirrors the same phases. Scaffold covers `create`, `scaffold enhance` and `scaffold upgrade`; develop covers `run`, `install` and `lint`; evaluate covers `eval run`, `eval generate`, `eval grade`, `eval dataset synthesize`, `eval compare`, `eval analyze`, `eval metric list` and `eval optimize`; deploy covers `deploy`, `publish gemini-enterprise`, `infra single-project` and `infra cicd`. The eval subcommands are the most interesting design choice: splitting inference (`eval generate`) from grading (`eval grade`) means you can regenerate traces without re-running the judge, and `eval compare` and `eval analyze` operate on saved result files rather than live runs.

Installing agents-cli and building a first agent

The README lists Python 3.11+, uv and Node.js as prerequisites. Installation is a single uvx invocation that installs the CLI and the skills together.

uvx google-agents-cli setup

If you only want the skills and intend to let your coding agent drive everything, the README gives a second path through the skills registry.

npx skills add google/agents-cli

After setup, open your coding agent and ask it to build something. The README's own example is deliberately small: "Use agents-cli to build a caveman-style agent that compresses verbose text into terse, technical grunts". The full walkthrough lives at the quickstart tutorial linked from the README. If you prefer to work without an assistant, the README states the CLI works standalone and points at `agents-cli scaffold` as the entry point; the command list also includes `agents-cli create <name>` for creating a new agent project and `agents-cli run "prompt"` for running the agent with a single prompt. Neither requires a cloud project, which matches the README's answer that local development can run on an AI Studio API key.

Where agents-cli stops being the right tool

The dependency on Google Cloud is the boundary, and it is not a soft one. The README is explicit: local work (`create`, `run`, `eval`) can run without a cloud project using an AI Studio API key, but "for deployment and cloud features, yes" you need Google Cloud. Deployment targets are named concretely: Agent Runtime, Cloud Run and GKE, with `infra single-project` and `infra cicd` provisioning the underlying infrastructure. If your agents run on AWS, on bare Kubernetes outside GKE, or on a self-hosted inference stack, the deploy and publish halves of this tool do not apply, and you are left with a scaffolding and eval CLI for a framework you may not be using. There is a second, quieter limitation. The README does not document rollback for `deploy` or for `scaffold upgrade`, and it does not describe what happens to an existing project when `scaffold enhance` adds CI/CD files that already exist. Those are the operations where a generated change can conflict with hand-written code, and the README's coverage of them is a one-line description each. Treat the first `enhance` on a long-lived repository as something to review as a diff, not to run unattended.

agents-cli compared with using ADK directly

The README anticipates this comparison and answers it in one line: ADK "is an agent framework", while agents-cli "gives your coding agent the skills and tools to build, evaluate, and deploy ADK agents end-to-end." The difference is not abstraction level, it is who writes the code. With ADK alone you read the framework docs and write the agent, the eval harness and the deployment config yourself. With agents-cli you describe the agent to your assistant, and the skills supply the ADK API surface, the eval methodology and the deployment shape. That is a real gain in speed and a real loss in transparency: the code that lands in your repository was written by a model working from skill files, so the quality of the output tracks the quality of those files and of your review. The same comparison applies to the coding assistants themselves. Antigravity CLI, Claude Code and Codex are named as compatible, and the README's FAQ states plainly that agents-cli is not an alternative to any of them. Choosing between those assistants is a separate decision from choosing this tool.

Maintenance, release cadence and the Apache-2.0 licence

The repository is not archived, and the last push was on 2026-09-03. Release history is dense and recent: v1.5.0 on 2026-09-01, v1.4.2 on 2026-08-28, v1.4.1 on 2026-08-24. That cadence matters for upgrade cost, because `agents-cli scaffold upgrade` exists specifically to move a project to a newer agents-cli version, and `agents-cli update` force-reinstalls skills to all IDEs. Both imply that the generated project layout and the installed skill files can drift apart from a pinned version, so a project created under v1.4.x and upgraded later is a supported path rather than a manual migration. The licence is Apache-2.0, which permits commercial use and modification and includes an explicit patent grant; it also requires that you preserve copyright and licence notices and state significant changes. That is a summary of the licence identifier in the repository, not legal advice, and the LICENSE file is the authority. One practical consequence of the licence combined with the architecture: the skills are files in the repository, so a team can read them, diff them between versions and vendor a pinned copy if the assistant's behaviour changes in a way they do not want.

Editorial conclusion

Adopt agents-cli if you are already building ADK agents and want your coding assistant to scaffold, evaluate and deploy them without you learning each Cloud CLI. Do not adopt it if you want a coding agent in itself, or if your target runtime is outside Google Cloud, since the deploy and publish commands point at Agent Runtime, Cloud Run, GKE and Gemini Enterprise. Before committing, verify three things on your own machine: that `uvx google-agents-cli setup` installs the skills into the assistant you actually use, that `agents-cli info` reports the project config and CLI version you expect, and that a local `agents-cli eval run` works with an AI Studio API key before any cloud project is involved.

Frequently asked questions

Official sources

1. google/agents-cli on GitHub
2. License: Apache-2.0
3. Project website
4. README
5. Releases