AtlasAgents

Local-first multi-agent coding system with Model Context Protocol (MCP) integration.

Architecture

AtlasAgents is a sophisticated orchestration system that manages AI coding agents through a controlled pipeline with human approval gates via Telegram. The system uses MCP for tool access and maintains strict workspace isolation.

Key Features

• Multi-Agent Pipeline: Planner → Spec-Writer → Coder → Reviewer workflow
• MCP Integration: Filesystem, Database, Git operations via Model Context Protocol
• Human-in-the-Loop: Telegram-based approval system for all mutations
• Local-First: All operations run locally with optional cloud model support
• Git-Based Safety: Shadow checkpoints and branch-based development
• Workspace Isolation: Each project runs in isolated workspace with root confinement

Installation

# Install in development mode
pip install -e .

# Install with dev dependencies
pip install -e ".[dev]"

Quick Start

# Initialize a new project
atlas init myproject

# Switch to project
atlas switch myproject

# Check status
atlas status

# Run planning stage
atlas run plan

Project Structure

/atlas/
  orchestrator/         # Core orchestration system
    core/              # State machine and orchestrator
    agents/            # Agent implementations
    mcp/               # MCP client and tool registry
    telegram/          # Telegram bot integration
    models/            # Database models
    utils/             # Utilities
  templates/           # Project templates
  workspaces/          # Project workspaces
  ~/.atlas/            # Global configuration
    settings.json      # Settings and model config
    history/           # Git checkpoints

Configuration

Settings are stored in ~/.atlas/settings.json. See tasks.md for detailed configuration options.

Development

# Run tests
pytest

# Format code
black orchestrator/
ruff check orchestrator/

# Type checking
mypy orchestrator/

License

MIT