AgentOS - Production AI Agent Runtime

AgentOS is a production-ready runtime for autonomous AI agents with built-in memory management, safe tool sandboxing, and multi-provider LLM support.

🎬 Demo

🚀 Quick Start

Installation

Then run the installer:

# Linux
python3 install_linux.py

# Windows
python install_windows.py

Basic Usage

1. Create an agent manifest (agent.yaml):

name: my_assistant
model_provider: github
model_version: openai/gpt-4o-mini
isolated: false

2. Run your agent:

agentos run agent.yaml --task "create a Python script that prints hello world"

3. Monitor running agents:

agentos ps

🏗️ Features

✅ Production Ready

• Comprehensive logging with structured output and per-agent log files
• Intelligent retry logic with exponential backoff for LLM API calls
• Process management with real-time monitoring and graceful shutdown
• Security controls blocking destructive commands and injection attacks
• Timeout protection preventing runaway processes
• Resource limits for memory, CPU, and execution steps

💬 Interactive Chat Mode

• Real-time conversations with AI using any LLM provider
• Rich terminal UI with markdown rendering and syntax highlighting
• Persistent chat history with SQLite backend and search functionality
• Conversation export to JSON, Markdown, or plain text formats
• Context preservation across sessions with configurable context window
• Customizable prompts and temperature settings
• Offline support with local Ollama models
• API-free options using GitHub or Ollama

🔒 Security First

• Command filtering blocks 20+ dangerous operations (rm, sudo, dd, etc.)
• Input validation prevents shell injection with pattern detection
• Path traversal protection blocks ../ and absolute path escapes
• Docker isolation (optional) with memory/CPU limits and network isolation
• Resource limits configurable per-agent (memory, CPU, timeout, steps)
• Security context for audit logging and tracking

🤖 Multi-LLM Support (6+ Providers)

• GitHub Models (default) - Free tier available
• OpenAI GPT-4o, GPT-4, GPT-3.5-turbo
• Anthropic Claude 3.5 Sonnet, Claude 3 Opus
• Google Gemini 2.0 Flash, 1.5 Pro
• Cohere Command R+, Command
• Ollama (local models) - No API key required

📊 Process Management

• Agent registry with SQLite backend
• Real-time process monitoring with CPU/memory tracking
• Status tracking (running, completed, failed, stopped)
• Log aggregation per agent with rotation support
• Graceful shutdown with signal handlers (SIGTERM/SIGINT)
• Agent lifecycle management with context managers

🔄 Retry Logic & Resilience

• Exponential backoff with configurable jitter
• Automatic retry for transient API failures
• Customizable retry strategies (aggressive, gentle, default)
• Per-provider retry configuration
• Circuit breaker patterns for failing services

📋 Commands

Run Agent

agentos run <manifest> --task "<task>" [--timeout 300] [--verbose]

Interactive Chat Mode ✨

Chat with any LLM provider in a conversational interface:

# Start chat with default OpenAI
agentos chat

# Use different providers
agentos chat --provider claude
agentos chat --provider gemini --temperature 0.3
agentos chat --provider ollama  # Local models, no API key needed

# Customize the experience
agentos chat --system-prompt "You are a Python expert"
agentos chat --provider openai --model gpt-4

In-chat commands: exit / quit (end), clear (history), help (commands), status (info)

See Chat Mode Guide for detailed usage.

List Agents

agentos ps

View Logs

agentos logs <agent_name> [--tail 50]

Stop Agent

agentos stop <agent_name>

Clean Up

agentos prune  # Remove stopped agents

📝 Agent Manifest

name: research_assistant
model_provider: github
model_version: openai/gpt-4o-mini
isolated: false

DESTRUCTIVE_COMMANDS:
  - rm
  - rmdir
  - sudo
  - dd
  - mkfs
  - format

Required Fields

• name: Agent identifier
• model_provider: LLM provider (github, openai, claude, gemini, cohere, ollama)
• model_version: Specific model to use

Optional Fields

• isolated: Enable Docker sandboxing (default: true)
• DESTRUCTIVE_COMMANDS: Custom list of blocked commands

🔧 Configuration

Environment Variables

Create .env file:

# API Keys (set as needed)
GIT_HUB_TOKEN=your_github_token
OPENAI_API_KEY=your_openai_key
CLAUDE_API_KEY=your_claude_key
GEMINI_API_KEY=your_gemini_key
COHERE_API_KEY=your_cohere_key

Logging

Logs are stored in ~/.agentos/logs/:

• agentos.log - Main system log
• <agent_name>_<id>.log - Per-agent execution logs

Database

Agent registry stored in ~/.agentos/runtime.db (SQLite)

🧰 MCP Tooling (Optional)

AgentOS can prefer MCP servers (Model Context Protocol) instead of emitting shell commands.

• Enable MCP in your manifest:

mcp:
  enabled: true
  servers:
    - name: local_tools
      kind: stdio
      command: my-mcp-server --stdio

• Install a Python MCP SDK (one of):

pip install mcp
# or install the official Model Context Protocol Python SDK if available

• Chat/Web will now prompt models to output MCP calls in a JSON block. AgentOS parses and executes those calls via the MCP client, with safe fallback to command extraction when no MCP calls are present.

🐳 Docker Support

Enable isolation for safe execution:

name: secure_agent
model_provider: github
model_version: openai/gpt-4o-mini
isolated: true

Requires Docker daemon running.

🛡️ Security Features

Command Filtering

Blocks dangerous commands automatically:

• File deletion: rm, rmdir, shred
• System modification: sudo, su, chown, chmod
• Disk operations: dd, mkfs, fdisk, format
• Process control: kill, killall, pkill
• Network: nc, netcat, wget, curl (to unknown hosts)

Input Validation

Prevents command injection attacks:

• Shell metacharacters: ;, &&, ||, |
• Command substitution: `, $()
• Variable expansion: $VAR, ${VAR}
• Path traversal: ../, absolute paths outside workspace

Resource Limits

Configure per-agent resource constraints:

resource_limits:
  max_steps: 50 # Maximum execution steps
  timeout: 300 # Timeout in seconds
  max_memory_mb: 512 # Memory limit (Docker only)
  max_cpu_percent: 50 # CPU limit (Docker only)

Security Context

Track and audit agent actions:

from agentos.core.security import SecurityContext, validate_command

with SecurityContext(agent_id="my_agent") as ctx:
    result = validate_command("ls -la")
    if result.is_safe:
        # Execute command
        pass
    # All actions logged automatically

🔄 Retry Configuration

Configure retry behavior for LLM API calls:

retry_config:
  max_retries: 3 # Maximum retry attempts
  initial_delay: 1.0 # Initial delay in seconds
  max_delay: 30.0 # Maximum delay cap
  exponential_base: 2.0 # Exponential backoff multiplier
  jitter: true # Add randomness to prevent thundering herd

Retry Strategies

from agentos.core.retry import DEFAULT_LLM_RETRY, AGGRESSIVE_RETRY, GENTLE_RETRY

# Default: 3 retries, 1-30s delay
config = DEFAULT_LLM_RETRY

# Aggressive: 5 retries, 0.5-60s delay (for critical operations)
config = AGGRESSIVE_RETRY

# Gentle: 2 retries, 2-10s delay (for user-facing features)
config = GENTLE_RETRY

💾 Chat History

Persistent chat history with SQLite backend:

from agentos.core.chat_history import ChatHistoryManager

# Initialize manager
history = ChatHistoryManager()

# Create conversation
conv_id = history.create_conversation(
    agent_id="assistant",
    title="Python Help Session"
)

# Add messages
history.add_message(conv_id, "user", "How do I read a file?")
history.add_message(conv_id, "assistant", "Use open() function...")

# Search history
results = history.search_messages("file", agent_id="assistant")

# Export conversation
history.export_conversation(conv_id, "chat.md", format="markdown")

🐳 Docker Sandbox

Enhanced Docker isolation for safe execution:

name: secure_agent
model_provider: github
model_version: openai/gpt-4o-mini
isolated: true

Advanced Docker Configuration

from agentos.core.docker_sandbox import DockerSandbox

sandbox = DockerSandbox(
    memory_limit="256m",      # Memory constraint
    cpu_quota=50000,          # CPU microseconds per period
    network_mode="none",      # No network access
    read_only=True,           # Read-only filesystem
    working_dir="/workspace"
)

result = sandbox.run_in_sandbox("python script.py")

Requires Docker daemon running.

📊 Process Monitoring

Real-time process monitoring and lifecycle management:

from agentos.core.process_manager import ProcessMonitor, AgentLifecycle

# Get singleton monitor
monitor = ProcessMonitor()

# Use lifecycle context manager
with AgentLifecycle("my_agent", task="Process data") as agent:
    # Agent is registered and tracked
    # CPU/memory monitored in real-time
    pass  # Do work
# Automatically cleaned up

# Query running agents
agents = monitor.get_running_agents()
for agent_id, info in agents.items():
    print(f"{agent_id}: {info['status']} - CPU: {info['cpu_percent']}%")

🛑 Graceful Shutdown

Signal handling for clean termination:

from agentos.core.shutdown import ShutdownManager, ShutdownContext

# Register cleanup callbacks
manager = ShutdownManager()
manager.register_callback(lambda: print("Cleaning up..."))

# Use context manager
with ShutdownContext():
    # Protected execution
    # SIGTERM/SIGINT handled gracefully
    pass

📊 Monitoring

Status Codes

• running: Agent is executing
• completed: Task finished successfully
• failed: Task failed with error
• stopped: Manually terminated

Exit Codes

• 0: Success
• 1: General error
• 124: Timeout
• 130: User interrupt (Ctrl+C)

🧩 Architecture

agentos/
├── agent/          # Agent execution and planning
├── cli/            # Command-line interface
├── core/           # Core utilities
│   ├── config.py       # Configuration management
│   ├── retry.py        # Retry logic with backoff
│   ├── security.py     # Security validation
│   ├── chat_history.py # Persistent chat storage
│   ├── shutdown.py     # Graceful shutdown
│   ├── docker_sandbox.py # Docker isolation
│   └── process_manager.py # Process monitoring
├── database/       # SQLite backend
├── llm/            # LLM provider integrations
├── mcp/            # Model Context Protocol
└── web/            # Web UI

🔄 Development

Local Setup

git clone https://github.com/agents-os/agentos
cd agentos
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Testing

python -m pytest tests/

Code Quality

black .
flake8 .

📄 License

MIT License - see LICENSE file.

🤝 Contributing

1. Fork the repository
2. Create feature branch (git checkout -b feature/amazing-feature)
3. Commit changes (git commit -m 'Add amazing feature')
4. Push to branch (git push origin feature/amazing-feature)
5. Open Pull Request

📞 Support

• Repository: https://github.com/agents-os/agentos
• Issues: GitHub Issues

---

AgentOS - Making AI agents production-ready, secure, and scalable.