anthropics/launch-your-agent: a Claude Code skill that takes you from idea to a live Managed Agent

The gap launch-your-agent fills between a CMA idea and a running agent

Claude Managed Agents give you primitives: an agent, an environment, a run, a deployment. The documentation is the source of truth for how those fit together, and the repository says so directly, pointing at platform.claude.com/docs/en/managed-agents/overview. What the docs do not do is sit with you and turn "I want an internal worker that triages inbound requests" into a concrete v0 with a definition of done.

That is the specific problem this repository targets. It is a Claude Code skill, not a library and not a service. You run Claude Code inside the cloned folder, type a slash command, and the skill conducts an interview, maps your answers to CMA primitives, stages a launch in your own account, grades the result against criteria you supplied, and, if the task recurs, puts it on a scheduled deployment. The audience is narrow and stated: a technical founder building an internal worker, a piece of their product, or a customer-facing agent.

The README is unusually candid about what this is not. It calls itself a reference implementation and states it is not maintained and not accepting contributions. It also warns that because the skill explains each step, it is more token-intensive than a purpose-built agent would be. Both of those are design decisions, not defects, and they should shape your decision before you clone anything.

Four phases, two skills, and a my-agent folder you keep

The mechanism is visible in the repository layout. The main skill lives in .claude/skills/launch-your-agent/ and is organized as four phases: interview, stage and launch, grade and iterate, run without you. Alongside those phases sit reference files: interview mapping, verified API call shapes, an examples bank, and an overview template. A companion skill in .claude/skills/wrap-up/ handles close-out and status checks.

The data flow is one-directional and file-based. Your interview answers map to CMA primitives through interview-to-config.md, and cma-primitives.md inventories those primitives and their limits from the public docs. The skill then writes everything into a my-agent/ folder in your working directory: the build sheet, the exact API payloads, a resumable launch script, an eval scaffold, an overview page, and a NEXT-DIRECTIONS.md that lays out v1 and v2. The launch script being resumable matters if a run fails partway; you are not starting the interview over.

The grading step is the part worth noting. The skill grades the launched agent against your own definition of done, which means the quality of that grade depends entirely on how precisely you answered the interview. Nothing in the repository supplies domain criteria for you. The examples-bank.md file offers sourced example agents and production proof points, but those are references, not a rubric.

Installing launch-your-agent and running your first agent build

There is nothing to install in the package-manager sense. The README's quickstart is a clone, a cd, and launching Claude Code. The skills under .claude/skills/ are picked up automatically when you run Claude Code inside the folder.

git clone https://github.com/anthropics/launch-your-agent.git
cd launch-your-agent
claude

Once Claude Code is open in that directory, you invoke the skill by typing the slash command on its own line.

/launch-your-agent

Before the flow can launch anything, you need an Anthropic API key for your own account. The README says you create it during the flow at platform.claude.com under API keys, and that it goes into a local .env file rather than into the chat. Runs cost cents, according to the README. That separation is the one security-relevant detail the documentation spells out, and it is worth respecting: do not paste the key into the conversation.

When the flow finishes, you have a live managed agent in your Console (agent, environment, graded run, and a scheduled deployment if the task recurs) plus the my-agent/ folder on disk. Later, from the same folder, /wrap-up refreshes the overview page, recaps every primitive you now own, and suggests the next one or two upgrades.

Where launch-your-agent is the wrong tool

The README's own disclaimer is the first limitation and the most consequential: this is a reference implementation, not maintained and not accepting contributions. If you file an issue expecting a fix, you have misread what you cloned. The last push to the repository was on 2026-07-07, and the project describes itself as unmaintained, so treat the shipped call shapes as a snapshot rather than a contract.

Second, the token cost. The skill explains each step as it goes, which the README says makes it more token-intensive than a purpose-built agent. If your goal is a cheap, repeatable build pipeline, a walkthrough that narrates itself is the wrong instrument. You are paying for the explanation.

Third, the API surface can drift. The repository states that the CMA documentation is the source of truth, which implicitly concedes that the verified call shapes bundled in the skill could fall behind the platform. Any launch failure that traces back to a changed payload shape is yours to diagnose against the live docs, not something the repository promises to patch.

Finally, this is not for non-technical users. The flow assumes Claude Code installed and signed in, an API key you manage yourself, and enough familiarity to read a launch script and an eval scaffold. A founder who wants an agent without touching payloads should look at a hosted product instead.

launch-your-agent versus hand-rolling against the CMA docs

The realistic alternative is not another skill. It is opening the Claude Managed Agents documentation and composing the primitives yourself, which is exactly what the repository says the docs are for.

The difference is in what you get and what you pay. Hand-rolling gives you full control over the payloads and no narration overhead, and it stays current because you are reading the live docs rather than a snapshot. It also means you design the interview-to-config mapping, the eval scaffold, the resumable launch script and the overview page yourself, or decide you do not need them.

launch-your-agent front-loads those artifacts. You get interview-to-config.md for the mapping, examples-bank.md for sourced reference agents, a build sheet and an overview page under ui/, and a NEXT-DIRECTIONS.md that separates v0 from v1 and v2. The trade is that you inherit a fixed four-phase flow and its explanation cost, and you accept that the bundled call shapes were verified at a point in time.

A second, narrower alternative is to read the repository without running it. cma-primitives.md and interview-to-config.md are standalone documents about CMA primitives and how interview answers map to them. If you only want the mental model, the clone-and-run step is optional.

Maintenance, licence and the cost of upgrading later

The repository carries an Apache-2.0 licence, stated in the README and present as a LICENSE file at the top level. Apache-2.0 permits commercial use, modification and redistribution, and it includes an explicit patent grant. That is a permissive arrangement, and the licence file itself is the authority on its terms; this is a description of what the repository states, not legal advice about your situation.

The maintenance picture is simpler and less comfortable. The README says the project is not maintained and not accepting contributions, and the last push was on 2026-07-07. Upgrades therefore do not arrive. If CMA primitives change, the skill does not follow automatically. Your upgrade path is manual: compare the verified call shapes in .claude/skills/launch-your-agent/ against the current CMA documentation, adjust, and re-run. The resumable launch script in my-agent/ is what makes that re-run tolerable rather than a full restart.

Budget for the token cost as an ongoing line rather than a one-time fee. Because the skill narrates each step, every iteration through the flow costs more than a purpose-built agent invocation would. The README frames this as the price of the explanation, and that framing is accurate.

Editorial conclusion

Adopt it if you are a technical founder who already has Claude Code and an Anthropic API key for your own account, and you want the API payloads, a resumable launch script and an eval scaffold written out rather than assembled from the CMA docs yourself. Skip it if you need a supported product with a maintenance commitment, if you want a purpose-built agent rather than a walkthrough, or if you are not prepared to pay the token cost the README admits to. Before you start, verify three things: that the CMA documentation at platform.claude.com still matches the verified call shapes shipped in .claude/skills/launch-your-agent/, that you have an API key scoped to your own account, and that you are willing to own the my-agent/ folder, because the repository states it is not maintained and not accepting contributions.

Frequently asked questions

Official sources

1. anthropics/launch-your-agent on GitHub
2. Issues
3. License: Apache-2.0
4. README