claude-dev-team

Claude Dev Team

A Claude Code plugin for multi-agent development workflows using Agent Teams.

Agents plan, implement, test, and review your code — coordinating through direct peer messaging.

Phase 1: Planning                   Phase 2: Development
┌──────────────────────────┐       ┌──────────────────────────┐
│  Lead (Orchestrator)     │       │  Lead (Orchestrator)     │
│  ├── Architect    [team] │       │  ├── Developer    [team] │
│  ├── Prod Manager [team] │       │  ├── Tester       [team] │
│  └── Researcher   [sub]  │       │  ├── Reviewer     [team] │
└──────────┬───────────────┘       │  └── Researcher   [sub]  │
           │                       └──────────┬───────────────┘
           └──── plan.md ─────────────────────┘

Install

claude plugin install claude-dev-team

Or from source:

claude --plugin-dir /path/to/claude-dev-team

Prerequisites

Agent Teams is experimental. The plugin checks this on session start and blocks if missing:

// .claude/settings.json or ~/.claude/settings.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

How It Works

Planning — Architect designs the system, PM validates requirements, Researcher fetches library docs. They debate directly until the design is solid. Output: plan.md.

Development — Developer implements in waves following the plan. Tester writes and runs tests, messaging the Developer directly on failures. Reviewer validates with /council (quick for routine, full review for security/architecture). All iterate via peer messaging until done.

The plan file bridges both phases — agents update it with status, logs, and changed files as they work.

Modes

Mode	Description
plan	Plan only: research, design, validate → plan.md
dev	Develop only: implement plan in waves → code + tests
full	Plan → user approval → develop → report
auto	Plan → develop → report (no approval gate)

Start here: Use full for most work. Use auto when you trust the planning to get it right without review.

Plugin Components

Skill

Single skill with four modes. The SKILL.md handles activation (what/when/roles/rules), while references/WORKFLOW.md is loaded on demand with detailed execution steps.

Agent

Researcher subagent — queries Context7 for library docs, searches the web for best practices, returns structured findings.

• MCP: Context7 is bundled via .mcp.json and auto-starts on plugin enable
• Memory: Project-scoped persistent memory avoids re-researching the same libraries across sessions
• Safety: Capped at 15 turns to prevent runaway loops

Hooks

SessionStart — validates CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is set. Blocks the session with an actionable error message if missing.

Agent Roles

Agent	Type	Model	Role
Architect	Teammate	Opus	System design, interfaces, component architecture
Product Manager	Teammate	Sonnet	Requirements validation, scope enforcement
Developer	Teammate	Opus	Implementation, bug fixes, iteration with tester/reviewer
Tester	Teammate	Sonnet	Test writing, execution, failure analysis
Reviewer	Teammate	Opus	Code quality, security, completeness — always validates with /council
Researcher	Subagent	Sonnet	Library docs via Context7, web search, best practices

Teammates are full Claude Code sessions that message each other directly — Architect debates with PM, Developer iterates with Tester and Reviewer.

Researcher is a subagent (no peer messaging). The Lead spawns it for doc lookups and relays findings to teammates.

Plugin Structure

claude-dev-team/
├── .claude-plugin/
│   └── plugin.json              # Plugin manifest
├── .mcp.json                    # Context7 MCP — auto-starts on enable
├── agents/
│   └── researcher.md            # Subagent with memory + Context7
├── hooks/
│   └── hooks.json               # SessionStart validation
├── scripts/
│   └── check-agent-teams.sh     # Prerequisite check
└── skills/
    └── claude-dev-team/
        ├── SKILL.md             # Activation: modes, roles, rules (76 lines)
        └── references/
            ├── WORKFLOW.md      # Execution: steps, prompts, templates (on demand)
            └── researcher-prompt.md  # Portable fallback for non-plugin use

Design Decisions

Why Agent Teams for most roles? Teammates can message each other directly. Architect↔PM debates and Dev↔Tester↔Reviewer iteration loops need back-and-forth — subagents can't do that.

Why subagent for Researcher? Research is a focused lookup: query docs, return results. No collaboration needed. The Lead relays findings.