v0.1 — local-only, runs on your Mac

my-team: Multi-Agent orchestration
for Claude Code.

Spin up a team of AI specialists that plan, code, test, review, and ship a PR — all from one command.

I built this because I wanted a more structured way to work with Claude Code — one that felt like collaborating with a coworker engineer instead of pair-programming with a single thread. It's now how I ship 95% of the time: hand a team a task, approve the plan, walk away to a PR.

View on GitHubQuickstart
core
4
optional
5
setup
1 cmd
output
1 PR
session.dispatch.loopwork in progress

The flow

One captain, a roster of specialists, one shared workspace.

Each session is a git worktree on its own branch. Four specialists are always on; five more are dispatched conditionally when their triggers fire. Agents communicate exclusively through files in .team/— srd, plan, tasks, journal, review. No hidden state, no chatter you can't inspect.

  1. The Captain

    The conversational anchor. Plans the work with you, dispatches specialists, ferries feedback between them, decides when the session is done.

Core specialists · 4

Dispatched on every session, in order, no triggers required. They turn a single command into a planned, tested, reviewed pull request.

  1. Scout

    core

    Read-only. Maps the codebase before any code is written, surfacing the files, conventions, and gotchas that will shape the plan.

  2. Engineer

    core

    Implements the plan task by task. Writes unit tests beside the code, commits to the session branch, leaves a journal entry for the next agent.

  3. Tester

    core

    Adds integration coverage, runs the full suite, files reproducible bug reports as severity-bucketed findings.

  4. Reviewer

    core

    Quality gate. Reads the diff, checks for blocking issues, leaves Suggestions, and either approves or sends fixes back to the engineer.

Optional specialists · 5

Dispatched only when the captain detects their trigger — a UI change for the designer, a stalled engineer for the debugger, sensitive files for the auditor, and so on.

  1. Debugger

    optional

    Investigation mode. When the engineer stalls on a failing test or unexplained behaviour, the debugger reproduces minimally, forms a hypothesis, and hands findings back via the journal.

  2. Designer

    optional

    Screenshot-driven visual critique for UI sessions. Boots a dev server, captures key views, calls out hierarchy and spacing issues, hands the engineer a revision list.

  3. Runner

    optional

    End-to-end behaviour check. Boots the actual feature like a real caller — dev server, CLI, function — inspects the response, flags any mismatch with the SRD or plan.

  4. Auditor

    optional

    Narrow security pass on auth, payments, PII, and migrations. Walks OWASP top-10 categories against sensitive files in the diff and appends findings under a Security audit section in review.md.

  5. Documenter

    optional

    Keeps README, CHANGELOG, AGENTS, and docs/ in sync after the engineer commits. Reads the diff for public-surface changes, updates the affected docs, commits separately.

How it works

One command. A worktree, a team, and a journal you can read.

~/code/wyvern
$ team new "Add user authentication"
  [my-team] creating worktree at ~/team/sessions/calm-river-12
  [my-team] spawning captain (claude)…
  captain › I'll scout the repo, draft a plan, and check in. Anything specific to call out?
$ > JWT, please. RS256.
  captain › Got it. Dispatching scout.
  [scout] reading 47 files in src/auth, src/middleware…
  captain › Plan ready. Approve to dispatch the team?
$ approve
  [engineer] feat(auth): add JWT signing service
  [tester] 14 specs passing, 1 flake fixed
  [reviewer] approved — 0 blocking, 2 suggestions
  [git] PR opened → github.com/you/repo/pull/482
.team/
shared state
.team/
├─ meta.json# session id, branch, source repo
├─ srd.md# session requirements doc — captain drafts with you
├─ plan.md# captain's plan — approved by you
├─ context.md# scout's codebase notes
├─ tasks.md# todo list for the team
├─ journal.md# append-only log of every step
├─ review.md# reviewer findings, severity-bucketed
├─ decisions.md# design choices logged inline
└─ state.json# phase, active specialist, errors
Every agent reads and writes here. You can too.

Why my-team

A whole team, on tap.

The shape of the work matches the shape of a real engineering team — split scope, lower cost, hands-off shipping.

Team-grade output

Specialists scout, build, test, and review in sequence — not one agent guessing.

Lighter tokens

Each agent sees only its slice, so total spend stays well below mega-prompt runs.

Walk away

Approve the plan once. Come back to a tested, reviewed, preview-deployed PR.

Architecture

Captain coordinates a team of specialists.

The captain is a long-running Claude Code process that dispatches sub-agents via the built-in Task tool and reads their artifacts back from the session's .team/ directory. Same token budget, same working tree. The infra strip up top — CLI, wrapper daemon, worktree — is how the captain gets booted.

team CLIcommander + chalkWrapper daemonHTTP · WS · node-pty~/team/sessions/<id>/git worktree + .team/INFRA · HOW THE CAPTAIN GETS BOOTEDspawnCaptainorchestrator · reads .team/claude (interactive)Task toolscout→ context.mdengineer×N parallel→ commitstester→ test runs · bugsreviewer→ review.mddebuggerconditional→ root-causedesignerconditional→ screenshotsrunnerconditional→ e2e transcriptauditorconditional→ security auditdocumenterconditional→ docs updates

Quickstart

Three commands. Then walk away.

  1. Clone
    git clone https://github.com/vik-srinivasan/my-team.git ~/.my-team

    Public repo. No auth needed.

  2. Install
    cd ~/.my-team && ./setup.sh

    Builds packages, links the team CLI globally. Needs Node 22 and pnpm 11.

  3. Ship
    team new "Add user authentication"

    From inside any git repo. Creates a worktree, spawns the captain, drops you into chat.

GitHubRead the full README

Get started

Install, run, inspect — or hand it to your agent.

From zero to a running team session, plus everything you need to come back later and pick up where the captain left off.

01 / SETUP

Install in two minutes

Prerequisites
  • Node 22+node -v
  • pnpm 11+pnpm -v
  • `claude` CLI installedclaude --version
  • GitHub CLI authenticatedgh auth status
Install
$ git clone https://github.com/vik-srinivasan/my-team.git ~/.my-team
$ cd ~/.my-team && ./setup.sh
$ team --help

If teamisn't found, add pnpm's global bin to PATH: echo 'export PATH="$(pnpm bin -g):$PATH"' >> ~/.zshrc