What is atl?
atl is a command-line tool that installs teams of AI agents into a project — the same way npm installs packages of JavaScript code, or brew installs Unix binaries.
The problem
Using Claude Code well requires configuration: agents, skills, and rules that shape how the model reasons about your codebase. You end up copying files between projects, forking someone else's setup, and slowly watching them drift apart. Every new project re-solves problems that a previous project already solved.
The answer
A team is a package of agents, skills, and rules, built around a particular kind of work. One team might be geared for a .NET + Flutter + React stack with a Docker-compose production layout. Another might be for a Next.js + Sanity + Vercel blog stack. A third for data pipelines with Airflow and dbt.
atl install some-team resolves the team against a GitHub-backed catalog, fetches its source, and copies its agents, skills, and rules into your current project's .claude/ directory. Claude Code sees the team the moment you open the editor.
When the team author ships a fix, you run atl update and every project that uses that team picks up the change. Your projects stop drifting.
Not a walled garden
Every team is just a public GitHub repository with a team.json file at the root. There's no central registry to submit to: tag a repo with the atl-team topic and it shows up in a generated catalog, where anyone can find it by handle and install it by <handle>/<name>. atl search queries that catalog; atl install resolves against it. The CLI is MIT-licensed Go. The team contract is documented here — see the team.json reference.
Who is this for?
- Developers who want a solid Claude Code setup without hand-rolling it for each project.
- Team leads who want to standardize how their company uses Claude across repos and onboard new engineers in minutes.
- Stack authors who want to publish opinionated agent teams the way framework authors publish CLIs today.
Where it stands
atl is v2 — a single monorepo (agentteamland/atl), currently in alpha. The install topology is project-local copies fetched from the catalog (no persistent on-disk clone cache — sources are fetched fresh and discarded after each install), the auto-update path runs through Claude Code SessionStart + UserPromptSubmit hooks, and the learning loop persists session knowledge: inline markers are queued durably, and the /drain skill folds each into the journal, wiki, and agent knowledge bases.
Two verified teams ship in the catalog today: software-project-team (13 agents — .NET API + Flutter + React + Docker stack) and design-system-team (2 agents + /dst-* skills for design-system + prototype tooling). The whole platform is MIT-licensed and open for contributions.
Next up: