Huddle

Decision-led development for teams coding with AI

Huddle is a Claude Code plugin for teams. One person hands out the tasks, every developer's AI session does one, and the people spend their time deciding and reviewing instead of typing.

Get started View on GitHub (opens in a new tab)

Five developers, five Claude Code sessions

AI tools made writing code cheap. They did not make a team faster on their own. When each person prompts the AI in their own way, each session understands the project differently, and the trouble shows up late, in review or in production.

  • Who does whatEvery task has its own Issue, its own branch and its own files, so two people do not collide.
  • Who decidesWhen a session meets a choice that is not its to make, it stops and asks the AI Manager, with options and a recommendation.
  • Nothing is droppedEverything an Issue asks for ends up done, or written down as not done, with the reason.

Every task shows where it is

An Issue moves along five stops. Labels on the Issue say where it is, and they update on their own. Two side roads exist because the AI is allowed to stop and ask.

The label flow of an Issue An Issue goes from ready to assigned, in progress, in review and merged. From in progress it can go to needs decision and come back once the AI Manager answers. From in review it can go to changes requested from AI Manager and come back to in progress. readyAI Manager writes it assignedGiven to a developer in progressA session works on it in reviewA pull request is open mergedAI Manager merges needs decision It asks, then goes on changes requested The AI Manager asks, the session fixes The label flow of an Issue An Issue goes from ready to assigned, in progress, in review and merged. From in progress it can go to needs decision and come back once the AI Manager answers. From in review it can go to changes requested from AI Manager and come back to in progress. readyWritten by the AI Manager assignedGiven to a developer in progressA session works on it needs decision in reviewA pull request is open changes requested mergedOnly the AI Manager merges
  1. ready: the AI Manager writes the Issue.
  2. assigned: the Issue is given to a developer.
  3. in progress: that developer's session works on it. Side road: if the session meets a choice that is not its to make, the Issue gets needs decision, the session asks the AI Manager, the AI Manager answers, and the Issue goes on.
  4. in review: a pull request is open. Side road: if the AI Manager asks for changes, the Issue gets changes requested from AI Manager and goes back to in progress, then to in review again.
  5. merged: only the AI Manager merges.

Details of every label: how it works.

How it works in four steps

  1. The AI Manager writes Issues

    Each task comes with the exact prompt to paste, the files it may touch and the tests that must pass.

  2. Each developer's session takes one

    The session reads the Issue, checks what it depends on, and works on its own branch.

  3. A pull request with tests

    First the failing test, then the code. The pull request lists the tests it ran, with real counts.

  4. The AI Manager reviews and merges

    They answer the questions, ask for changes or merge. Only they merge, and only they touch production.

Two roles, one flow

Huddle splits the work between a person who decides and a session that does. Each one knows what is theirs.

The AI Manager

  • Writes the Issues, with the prompt, the allowed files and the tests.
  • Answers when a session stops to ask for a decision.
  • Reviews every pull request, and is the only one who merges.

The developer's session

  • Takes one Issue and works on its own branch.
  • Writes the failing test first, then the code.
  • Stops and asks when a choice is not its to make, then opens a pull request.

GitHub and GitLab

Huddle works on github.com and on GitHub Enterprise Server, and on gitlab.com and on a company's own GitLab server. Huddle uses the login of gh or glab and never reads or stores a token.

The same commands everywhere.

GitHub

A workflow keeps the labels up to date whenever an Issue or a pull request changes.

GitLab

A scheduled pipeline updates the labels every ten minutes. The AI Manager sets up the schedule and one token, once.

Install it in two commands

Every developer installs the plugin once, on their own machine. You need Claude Code, Node 22 or newer, git, and gh or glab. Run these in a terminal, then open a new Claude Code session:

Terminal 
claude plugin marketplace add SrProfessional/Huddle
claude plugin install huddle@huddle

Then, in your repo, the AI Manager runs /huddle-setup. The get started guide has every step. Huddle has no runtime dependencies and nothing to npm install.

Questions people ask

What is Huddle?

A team method and a Claude Code plugin. One person, the AI Manager, hands out tasks as Issues. Each developer's Claude Code session works on one task in its own branch, asks when a decision is not its to make, and opens a pull request for review.

How do teams use Claude Code together?

By giving every session the same context and the same rules. The Issue holds the goal, the allowed files and the tests, and the repo's CLAUDE.md holds the rules, so sessions started by different people behave the same way.

Does it work with GitLab?

Yes. It works with GitHub and GitLab, including a self-managed GitLab server. On GitLab the labels update on a pipeline schedule instead of reacting at once.

What does decision-led development mean?

The AI writes, tests and asks. The person decides. Our job is no longer typing code but making good decisions, and a session that cannot go on without one stops and asks instead of guessing.

Does Huddle need a server, an account or a token?

No server and no account. It runs on your machine with Node and the GitHub or GitLab command line you already log in to. The only token is a GitLab CI variable the AI Manager sets up for the label job.

Is it free?

Yes. Huddle is open source under the MIT license, and it has no dependencies and no trackers.

Ready when your team is

Two commands to install, one setup in your repo, and the next Issue is already on its way.

Get started