Claude Agents - AI-Powered SDLC Agent System

13 auto-activating AI agents for every phase of software development. Just describe what you want to build — the right specialists engage automatically.

Works with Claude Code (Anthropic's CLI). No plugins, no configuration, no commands to memorize.

Requirements: macOS or Linux with bash and git; jq and npx are optional (used for the statusline and MCP examples). See INSTALL.md for updating, cross-machine migration, and the autonomous self-heal setup.

30-Second Setup

git clone https://github.com/pfangueiro/claude-code-agents.git
cd claude-code-agents
./install.sh

Installs once, user-global to ~/.claude. Now every project on your machine has the agents — open any project with Claude Code and talk naturally:

"Build a REST API with JWT authentication"
"This query is running slow"
"Check for security vulnerabilities"
"Deploy to production"

Agents activate automatically based on your words.

What You Get

Available Agents

Cost optimization: Haiku for docs (95% cheaper), Sonnet for dev work, Opus for security/emergencies.

How Agents Collaborate

Agents hand off work automatically. Example: "Build a user registration system"

graph LR
    A[architecture-planner] --> B[frontend-specialist]
    A --> C[api-backend]
    C --> D[database-architect]
    B --> E[test-automation]
    C --> E
    E --> F[code-quality]
    F --> G[security-auditor]
    G --> H[devops-automation]

See .claude/lib/agent-coordination.md for the formal handoff protocol.

Installation

One install, user-global. Claude Code natively loads ~/.claude/{agents,skills,commands,rules}, so a single install makes the agents available in every project on your machine — no per-project setup.

Verify Installation

./validate.sh
# All validations passed! (full + --quick modes)

Installed components live under ~/.claude/ — verify with ls ~/.claude/agents ~/.claude/skills ~/.claude/commands ~/.claude/rules.

Slash Commands

Project Workflow

Developer Workflow (new in v2.5.0)

Skills System

Skills provide domain knowledge that agents apply. 28 included:

Create Your Own Skill

python3 .claude/skills/skill-creator/scripts/init_skill.py my-skill --path .claude/skills

MCP Integration

4 Model Context Protocol servers, provided as an example config (.mcp.json.example) — install does not register them; add the ones you want to Claude Code with the commands below:

Manual MCP Setup

claude mcp add context7 -- npx @upstash/context7-mcp@latest
claude mcp add sequential-thinking-server -- npx @modelcontextprotocol/server-sequential-thinking
claude mcp add playwright -- npx @executeautomation/playwright-mcp-server
claude mcp add github -- npx -y @modelcontextprotocol/server-github

Auto-Enforced Rules

Six rule files in ~/.claude/rules/ are automatically loaded by Claude Code in every session:

• security.md — No secrets in commits, parameterized queries, input validation, security headers, least privilege
• code-quality.md — No dead code, single responsibility, early returns, explicit error handling, descriptive naming
• fix-quality.md — Root cause analysis before fixing, never suppress errors, minimal changes, test-driven fixing
• verification.md — Verify after every implementation, run tests, build to catch errors, test-driven bug fixes
• framework-integrity.md — Hooks must enforce invariants, state reconciles automatically, snapshots have documented restore paths
• anti-anchoring.md — Name the obvious pick + a non-obvious alternative + a trap on open-ended judgment calls; escalate consequential open decisions to /diverge

Hooks

10 hook events across 9 command hooks, installed globally to ~/.claude/hooks/:

Phase 2 reference configs (opt-in, not enabled by default):

• smart-file-guard.json — prompt hook: LLM-based file protection
• pre-commit-review.json — agent hook: automated code review before commit

Architecture

Installed once, user-global under ~/.claude/ — Claude Code loads it in every project:

~/.claude/
├── agents/          # 13 auto-activating SDLC agents
├── commands/        # 13 slash commands
├── skills/          # 28 modular knowledge packages
├── rules/           # 6 auto-enforced rule sets
├── lib/             # Templates, patterns, coordination protocol
├── hooks/           # 9 command hooks + 2 reference configs
├── daemon/          # launchd watchdog (hourly validate + snapshots)
├── analytics/       # Observability dashboard + ingested session logs
├── statusline.sh    # Rich status bar — ends line 1 with a framework-status glyph (✓ healthy / ⟳ self-healing / ⚠ stalled-or-degraded)
└── settings.json    # 10 hook events, permissions, model config

See EXTENSIBILITY.md for the complete guide on Skills, MCP, Slash Commands, and Subagents.

Usage Examples

Building a feature:

You: "Build a shopping cart with Stripe integration"
  → architecture-planner designs the system
  → api-backend implements payment logic
  → database-architect creates order schema
  → security-auditor validates payment security
  → test-automation generates tests
  → documentation-maintainer documents the API

Performance issue:

You: "The product search is running really slow"
  → performance-optimizer profiles the code
  → database-architect optimizes queries
  → code-quality suggests improvements

Production emergency:

You: "CRITICAL: Production API is returning 500 errors!"
  → incident-commander takes charge (Opus)
  → security-auditor checks for breaches
  → devops-automation prepares rollback

Observability Dashboard

Built-in analytics dashboard that reads Claude Code's native JSONL session logs — zero cloud dependencies, Python stdlib only, single SQLite database.

# Quick start (after install)
claude-obs

# Or run manually
python3 ~/.claude/analytics/collector.py    # Ingest JSONL → SQLite
python3 ~/.claude/analytics/server.py --open # Serve dashboard at localhost:3141

Dashboard panels:

• Summary cards — total cost, projects, sessions, agent activations, input/output tokens
• Daily cost & sessions — dual-axis bar+line chart with configurable time ranges (7d/14d/30d/All)
• Cost by project — horizontal bar chart ranking projects by spend
• Agent activations — which agents are used most and across how many projects
• Model distribution — doughnut chart showing Opus/Sonnet/Haiku cost split
• Top projects by tokens — token consumption per project
• Sessions table — recent sessions with project, duration, cost, model, and agents used

Architecture:

• Data source: ~/.claude/projects/**/*.jsonl (Claude Code writes these automatically)
• Collector (collector.py): incremental JSONL scanner with per-file watermarks, per-model cost estimation (Opus/Sonnet/Haiku pricing)
• Storage: SQLite with WAL mode at ~/.claude/analytics/claude-obs.db
• Server (server.py): 6 JSON API endpoints + static file server on localhost:3141
• Dashboard (dashboard.html): single-file dark-themed UI with Chart.js

CLI options:

python3 collector.py --full          # Re-ingest everything (ignore watermarks)
python3 collector.py --db /path/to.db # Custom database path
python3 server.py --port 8080        # Custom port
python3 server.py --open             # Auto-open browser

Installed automatically by ./install.sh. Data stays entirely local.

Contributing

1. Fork the repository
2. Create your feature branch (git checkout -b feature/my-feature)
3. Run ./validate.sh to verify everything passes
4. Commit your changes (git commit -m 'feat: add my feature')
5. Push and open a Pull Request

Roadmap

• VS Code extension
• Observability dashboard
• Lifecycle hooks (agent tracking, permission auditing, session lifecycle)
• Additional specialized agents
• Multi-language support
• Team collaboration features
• Custom agent marketplace

License

MIT License - see LICENSE for details.

Support

• GitHub Issues
• GitHub Discussions
• Update: ./install.sh --update