Sage Team v3.7

AI-Powered Autonomous Software Company — 11 agents execute real code via Claude Code, with a PixiJS isometric office visualization and native Claude Code MCP integration.

Sage Team runs a complete AI software company. A CEO agent decomposes your goal into a sprint, then 11 specialized agents execute tasks in parallel using Claude Code subprocesses — writing real code, running tests, and creating PRs. Watch it all happen in a PixiJS isometric office in your browser.

What Makes This Different

• Real Execution — Agents run Claude Code subprocesses that write, test, and commit actual code
• Hybrid Engine — CEO plans via Claude API (fast/cheap), agents execute via Claude Code (powerful)
• SQLite Persistence — Full state survives restarts. Resume any session.
• Skill Injection — Agent prompts include real skill protocols from Superpowers and Antigravity
• Isometric Office — PixiJS 2.5D office with 10 rooms, pixel-art agents that walk around, particle effects
• Visual Scrum Board — Live task tracking with To Do / In Progress / Done columns
• PR Workflow — Agents create branches, you review and approve/reject PRs in the UI

Quick Start

Option A: Inside Claude Code (recommended)

npm install -g sage-team
cd your-project
claude

Inside Claude, just talk naturally:

"Initialize sage team here" "Start the team with the goal: Build a REST API for a todo app" "How's the sprint going?" "Approve the PR"

Claude Code auto-discovers Sage Team via MCP. Set up the .mcp.json in your project:

{
  "mcpServers": {
    "sage-team": {
      "command": "sage-team-mcp",
      "args": []
    }
  }
}

Option B: Standalone CLI

# Install globally
npm install -g sage-team

# Set your API key
export ANTHROPIC_API_KEY=your-key-here

# Initialize project
sage-team init

# Launch the office
sage-team start

# Or start with a goal
sage-team start --goal "Build a REST API for a todo app with authentication"

The Team

| Agent | Role | Room | Focus | |-------|------|------|-------| | 👑 Sage | CEO | CEO Office | Goal decomposition, sprint planning, delegation | | 🔬 Nova | CTO | CTO Office | Architecture review, technical decisions | | 🏛️ Aria | Architect | Arch Lab | System design, API patterns | | ⚡ Dex | Senior Dev | Dev Bullpen | Core implementation, TDD | | 🌊 Flux | Full Stack Dev | Dev Bullpen | Frontend + backend, end-to-end features | | 🔍 Quinn | QA Lead | QA Lab | Testing, quality gates | | 🚀 Gage | DevOps | QA Lab | CI/CD, deployment, infrastructure | | 📋 Morgan | Product Manager | Design Studio | Requirements, prioritization | | 🎨 Uma | UX Designer | Design Studio | UI/UX design, accessibility | | 🌀 River | Scrum Master | Lounge | Process, blocker removal | | 📊 Atlas | Data Engineer | Data Lab | Database, data pipelines |

Browser UI

┌──────────────────────────────────────────────────────┐
│ 👑Sage 🔬Nova 🏛️Aria ⚡Dex 🌊Flux 🔍Quinn ...       │  Agent bar
├───────────┬──────────────────────┬───────────────────┤
│ SCRUM     │                      │ ACTIVITY          │
│ BOARD     │   ISOMETRIC OFFICE   │ FEED              │
│           │   (PixiJS canvas)    │                   │
│ ● Doing   │   Agents walk around │ sage: planning..  │
│   Dex     │   rooms, sit at      │ dex: coding...    │
│   coding  │   desks, visit       │ quinn: testing..  │
│           │   each other         │                   │
│ ● To Do   │                      │                   │
│ ● Done    │                      │                   │
│           │                      │                   │
│ Team      │                      │                   │
│ 👑🔬🏛️⚡🌊│                      │                   │
├───────────┴──────────────────────┴───────────────────┤
│ [Enter a goal for your team...               ] [Go]  │
└──────────────────────────────────────────────────────┘

• Left — Scrum Board: task columns (In Progress, To Do, Blocked, Done) + team avatars
• Center — Isometric office: pixel-art agents wander between 10 rooms with solid walls and doorways
• Right — Live activity feed: timestamped events from all agents as they work
• Bottom — Goal input + pending PR badges

MCP Tools (inside Claude Code)

When using Sage Team inside Claude Code, these tools are available:

| Tool | Description | |------|-------------| | sage_team_init | Initialize Sage Team in current project | | sage_team_list | Show all 11 agents with roles and skills | | sage_team_start | Start a session with a goal | | sage_team_status | Show sprint progress, tasks, and agent statuses | | sage_team_doctor | Check system health (Node, Git, Claude Code, API key) | | sage_team_config | View or update configuration | | sage_team_approve_pr | Approve and merge a pull request | | sage_team_reject_pr | Reject a PR with feedback for rework |

CLI Commands

sage-team init                    # Initialize .sage-team/ directory
sage-team start                   # Launch office + browser UI
sage-team start --goal "..."      # Start with an initial goal
sage-team start --no-browser      # Skip opening browser
sage-team start --port 8080       # Custom port (default: 3000)
sage-team resume                  # Resume latest session
sage-team resume --session <id>   # Resume specific session
sage-team config --show           # Show configuration
sage-team config --api-key <key>  # Set API key
sage-team config --model <model>  # Set Claude model
sage-team team                    # Show team roster
sage-team status                  # Show session status
sage-team doctor                  # Check system health
sage-team logs                    # Tail event log
sage-team logs --agent dex -n 50  # Filter by agent

How It Works

You submit a goal
        │
        ▼
┌─────────────┐     Claude API
│  CEO (Sage) │────────────────► Decompose into sprint + tasks
└─────────────┘                  with dependencies & assignments
        │
        ▼
┌─────────────┐     Claude Code
│  Dispatcher │────────────────► Spawn subprocesses (max N concurrent)
└─────────────┘                  Each agent gets: identity + skills + task
        │
        ▼
┌─────────────┐
│   Agents    │────────────────► Write code, run tests, commit
│  (parallel) │                  Stream status back via JSONL
└─────────────┘
        │
        ▼
┌─────────────┐
│ PR Manager  │────────────────► Create branch, squash merge on approval
└─────────────┘
        │
        ▼
┌─────────────┐     WebSocket
│   Browser   │◄───────────────── Real-time events, agent positions,
│   (PixiJS)  │                   scrum board, particle effects
└─────────────┘

Execution Modes

• Sandbox (default) — Each agent works in a git worktree. PRs require your approval.
• Direct — Agents commit directly to the working branch. Fast but risky.
• Supervised — Like sandbox, but agents pause and ask before critical operations.

Architecture

sage-team/
├── src/
│   ├── types/              # TypeScript interfaces (Agent, Task, Sprint, etc.)
│   ├── state/
│   │   ├── database.ts     # SQLite with WAL mode, 9 tables
│   │   └── repositories/   # 7 repos (sessions, agents, tasks, sprints, etc.)
│   ├── agents/
│   │   ├── personas.ts     # 11 agent definitions with skills & colors
│   │   └── skills-map.ts   # Agent → skill ID mapping
│   ├── skills/
│   │   └── loader.ts       # Loads Superpowers/Antigravity skill protocols
│   ├── engine/
│   │   ├── ceo-brain.ts    # Goal decomposition via Claude API
│   │   ├── dispatcher.ts   # Claude Code subprocess management
│   │   ├── orchestrator.ts # Central coordinator (tick loop every 5s)
│   │   ├── pr-manager.ts   # Git worktree + branch + merge
│   │   └── prompt-builder.ts # Identity + skills + task prompt assembly
│   ├── cli/
│   │   ├── index.ts        # Commander program with 8 commands
│   │   └── commands/       # init, start, resume, config, team, status, doctor, logs
│   ├── mcp/
│   │   └── server.ts       # MCP server for Claude Code integration
│   ├── server/
│   │   ├── index.ts        # Express + WebSocket server
│   │   └── routes/api.ts   # REST API endpoints
│   └── web/                # Frontend (separate Vite project)
│       └── src/
│           ├── canvas/     # PixiJS isometric office, pixel-art agents, effects
│           ├── hooks/      # useWebSocket, useOffice, usePixiOffice (wandering sim)
│           ├── panels/     # AgentBar, TaskBoard (Scrum), ChatPanel, PRReview, GoalInput
│           └── store.ts    # Zustand state management
├── docs/
│   └── USAGE-GUIDE.md     # Step-by-step usage guide
├── vitest.config.ts        # Test configuration
└── package.json

Configuration

# API key (required)
export ANTHROPIC_API_KEY=your-key-here

# Or save to config
sage-team config --api-key your-key-here

# Customize
sage-team config --company-name "My AI Company"
sage-team config --model claude-sonnet-4-20250514
sage-team config --max-agents 5
sage-team config --autonomy sandbox

Skill Sources

Agent prompts are injected with real skill protocol content from:

• Superpowers — TDD, brainstorming, debugging, code review, verification
• Antigravity — TypeScript, React, security, DevOps, databases, and more

Skills are loaded at runtime from your local installations and respect a 30k token budget per agent prompt.

Prerequisites

• Node.js >= 18
• Git
• Claude Code CLI installed (npm i -g @anthropic-ai/claude-code)
• Anthropic API key

License

MIT