The agent team
we run every day.
Now it's yours too.
8 production-tested Claude Code subagents + the orchestration rules that keep them from stepping on each other. Free, MIT-licensed, no signup.
Why We Built This
We use Claude Code for real client work every day, and one prompt trying to be a backend dev, a frontend dev, and a sysadmin at once falls apart fast. So we split the work into specialists, gave each one exactly one job and one doc file to maintain, and wrote down the rules that keep them from colliding. That system is what's in this repo.
Nothing here is theoretical. This is the exact agent pack running in
our own ~/.claude/agents/ right now, scrubbed of
client-specific details and published as
github.com/ironmindai/claude-dev-team.
MIT licensed. Take it, fork it, rip out what you don't need.
The Roster
backend-coder
Server-side code — Python, Node, Go, Rust. APIs, business logic, DB interactions. Maintains docs/backend-routes.md.
frontend-brand-guardian
All frontend code. Enforces — and bootstraps if missing — a docs/brandbook.md so every page stays on-brand. (This page was written by it.)
frontend-backend-sync-reviewer
Runs after backend/frontend work to verify every frontend API call has a matching backend route. Catches breakage before runtime.
system-devops-admin
The only agent allowed to touch the system — packages, nginx, SSL, systemd, ports, venvs. Maintains docs/system-devops-admin.md.
database-schema-manager
Sole authority for schema changes. Every migration is reversible Alembic — no ad hoc ALTER TABLE from other agents.
scheduled-tasks-coder
Cron jobs, systemd timers, background workers. Maintains docs/scheduled-tasks.md so nobody double-schedules a job.
llm-workflow-architect
Designs multi-step LLM pipelines with structured outputs, logging, and cost control — for when a single prompt isn't enough.
headless-http-explorer
Maps a web app's real HTTP/API surface by observing network traffic. Headless only — no scraping, no UI clicking.
ui-browser-debugger (optional)
Visual UI debugging with a real browser. Requires a browser automation setup — only wire it in if you have one.
Three Rules, Not a Pile of Prompts
Flat call tree
Only the orchestrator spawns agents. Specialists can't spawn each other — it's not in their tool list, so it's enforced by the harness, not prompt hope. Every run stays auditable.
Docs as shared memory
Each specialist owns one file in docs/, reads it before acting, updates it after real changes. That's how agents coordinate without talking directly.
Decisions made once
Ports and venvs get allocated up front by the orchestrator, never ad hoc mid-task. Parallel agents get a pre-decided shared contract instead of guessing at each other.
Install It
The fastest way in: open Claude Code, point it at the repo, and say
"install this." The repo ships an INSTALL.md written
specifically for an AI agent to follow step by step — clone, copy agents
into ~/.claude/agents/, merge the routing rules into your
CLAUDE.md, done. No manual file surgery required.
Prefer to read it yourself first? The README walks through the same
steps manually, plus a project-level option if you want to trial it
scoped to one repo before promoting it to ~/.claude/.