Bro — AI skill for Claude Code

bro

[!WARNING] **Early development.** Interfaces move without deprecation cycles and `master` carries breaking changes. Provided as-is, without warranty or support; use it at your own risk.

**Declare an agent once. Run it on any harness[^harness] and any LLM[^llm]. Know exactly what it can reach.**

A bro is an agent persona declared in one place — what it knows, which tools it holds, which secrets it may read, whom it may delegate to — with nothing about it wired anywhere else. The same declaration runs as a Claude Code session or under bro's own agent loop, one-shot or interactive, on whichever model you name at launch. Wherever it runs, it runs inside a boundary that holds exactly what the declaration says: a fresh container, a clone, the credentials it needs and no other, the bros it may summon and no other. The rest of this page shows all of that on one made-up bro.

Getting started

The way in is [`BOOTSTRAP.md`](BOOTSTRAP.md), an executable checklist for an agent rather than a page for you. On a macOS or Ubuntu host with Git, Python 3.12 or newer, [uv](https://docs.astral.sh/uv/), Docker, and Claude Code, open an agent session in the repository you want to adopt bro in and point it at the file; any harness that can read and edit the repository will do. With Claude Code:

cd ~/acme
claude 'fetch https://raw.githubusercontent.com/dzhioev/bro/master/BOOTSTRAP.md and follow its instructions'

The session pins the framework into the project from this repository at one commit, declares a first developer bro, lets you choose which harness runs launches and summons by default, wires the host's credentials, and commits the setup; then it proves the setup with an isolated run under each harness whose credential is wired and hands you a `dive-in` to continue from, showing every write before it makes it.

To work on the framework itself, clone this repository and run `./setup.sh`.

Meet triage, an example

Everything from here on is one