CodeCrew

A virtual dev crew for Claude Code — 11 specialized AI agents that decompose, implement, review, and ship your tasks.

Instead of one AI doing everything, CodeCrew breaks work into subtasks and routes each to the right specialist on the right model. A 3-layer codebase index keeps every agent focused on exactly the lines they need, achieving 5-10x token savings per interaction.

Quick Install

# Add marketplace and install
/plugin marketplace add d3x293/code-crew
/plugin install codecrew@codecrew

# Initialize on your project
/crew init

# Run your first task
/crew task "fix the authentication bug in the login flow"

Why CodeCrew?

• Specialization over generalism — A debugger investigates, a senior dev fixes, a reviewer checks. Each agent is purpose-built for its role, producing better results than a single generalist prompt.
• 5-10x token savings — The 3-layer codebase index means agents read ~300-500 tokens to find and edit a function, instead of ~2000-10000+ tokens reading entire files.
• Parallel execution — Independent subtasks run simultaneously. Implementation and test writing happen in parallel, cutting task completion time.
• 125+ skills auto-injected — CodeCrew scans your installed Claude Code plugins, finds the best-fit skills for each task, and injects them into agent prompts with strict token budgets.
• Intelligent model routing — Opus handles complex architecture decisions. Sonnet handles implementation and review. Haiku handles simple fixes and docs. The right model for the right job.
• Domain-aware setup — Detects your project type (trading bot, e-commerce, SaaS, etc.) and suggests specialized custom agents tailored to your domain.

Works Best With

CodeCrew automatically discovers and uses skills from other installed Claude Code plugins. It works standalone, but unlocks its full potential with:

everything-claude-code — Provides 125+ development skills (react-patterns, tdd-workflow, security-scanning, architecture-decision-records, and more). CodeCrew's skill injector selects the most relevant skills for each agent based on task type, project stack, and agent role.

superpowers — Extends agent capabilities with advanced workflows like systematic debugging, test-driven development, code review protocols, and parallel agent dispatch.

How skill injection works:

1. During /crew init, CodeCrew catalogs all skills available from installed plugins
2. When a task runs, the skill injector matches skills to the task type + project stack + agent role
3. Skills are condensed to key guidance (~150 tokens each) and injected with strict token budgets
4. Opus agents get up to 6 skills (3000 tokens), Sonnet gets 4 (2000 tokens), Haiku gets 2 (800 tokens)

This prevents context bloat while giving every agent the exact guidance it needs.

How It Works