Atlas — AI Agent Harness for VS Code

An autonomous multi-agent AI engineering operating system embedded inside VS Code. Atlas coordinates multiple specialized AI agents to understand your codebase, plan complex engineering tasks, execute them transparently, and keep you in control.

Zero cost. Zero API keys required. Works offline.

Features

• Multi-Agent Pipeline — 14 specialized agents (Planner, Architect, Coder, Reviewer, Debugger, Tester, Git, Security, Performance, and more) that collaborate to complete tasks
• Inline Edit (Ctrl+Alt+K) — Select code, describe the change, Atlas rewrites it inline
• Chat Panel — Full conversational AI assistant with code block detection and one-click apply
• Agent Orchestrator — Spawn, kill, pause, resume agents; assign work, track progress, handle failures
• Workspace Analysis — Automatically understands project type, languages, frameworks, dependencies, build systems, and architecture
• Memory System — 7 memory stores (working, conversation, workspace, long-term, errors, project knowledge, decisions)
• Codebase Index — Symbol-level workspace search and retrieval for smart context
• Tool System — 19 built-in tools (file ops, git, terminal, code analysis, build/test runner) with retry, permissions, and rollback support
• Observability — Full logging, tracing, metrics collection, and execution replay
• Safety — Configurable approval mode (always prompt, dangerous only, never)
• Provider-Agnostic — Supports Ollama (local), OpenAI, Anthropic, Google, OpenRouter, LM Studio
• Extensible — Plugin system for custom agents, tools, providers, and memory backends

Quick Start

1. Open this folder in VS Code
2. Press F5 to launch the Extension Development Host
3. Atlas will auto-detect Ollama (if running) or guide you through setup

No API key? No problem. Atlas works with:

• Ollama (local, free, offline) — install from ollama.ai, launch it, Atlas auto-detects it
• OpenRouter free tier — sign up at openrouter.ai for free credits
• Google Gemini free tier — free API key from aistudio.google.com

Keybindings

Local AI Setup (Ollama)

Atlas works with Ollama completely offline and free.

1. Download Ollama and launch it
2. Atlas auto-detects it on startup and walks you through pulling a model
3. Start using Atlas with zero configuration

Recommended local models:

• llama3.2:1b — fastest, ~0.6GB RAM
• llama3.2:3b — balanced, ~2GB RAM
• llama3.1:8b — most capable, ~4.7GB RAM
• deepseek-coder:6.7b — specialized for code
• qwen2.5:7b-coder — specialized for code

Architecture

src/
├── extension.ts              # Entry point
├── activation/               # Extension bootstrapping + setup wizard
├── agents/                   # 14 specialized AI agents
│   ├── BaseAgent.ts
│   ├── PlannerAgent.ts       # Breaks goals into tasks
│   ├── ArchitectAgent.ts     # Designs architecture
│   ├── CoderAgent.ts         # Writes code
│   ├── ReviewerAgent.ts      # Reviews code quality
│   ├── DebuggerAgent.ts      # Investigates failures
│   ├── TesterAgent.ts        # Writes tests
│   ├── GitAgent.ts           # Manages commits
│   ├── SecurityAgent.ts      # Detects vulnerabilities
│   ├── PerformanceAgent.ts   # Profiles performance
│   ├── DependencyAgent.ts    # Updates packages
│   ├── RefactorerAgent.ts    # Improves code quality
│   ├── DocumentationAgent.ts # Generates documentation
│   ├── ResearchAgent.ts      # Searches codebase
│   └── TerminalAgent.ts      # Executes commands
├── orchestrator/             # Agent lifecycle & task management
├── collaboration/            # Event bus, workflows, arbitration
├── workspace/                # Analysis, semantic map, codebase index
├── memory/                   # 7 memory stores
├── context/                  # Intelligent context retrieval
├── tools/                    # 19 tools with retry & permissions
├── providers/                # LLM providers (OpenAI, Anthropic, Ollama, etc.)
├── safety/                   # Approval system
├── observability/            # Logging, tracing, metrics, replay
├── task/                     # Task planning & execution
├── ui/                       # Dashboard, chat, tree views, status bar
├── settings/                 # Configuration manager
└── extensibility/            # Plugin system & public API

Pipeline Flow

User Request
    ↓
  Planner — breaks goal into tasks
    ↓
  Architect — understands project structure
    ↓
  Coder(s) — implement in parallel
    ↓
  Reviewer — reviews code quality
    ↓
  Tester — writes and runs tests
    ↓
  Git Agent — commits changes
    ↓
  Report

Configuration

Settings available under atlas.*:

Free vs Paid Providers

License

MIT