swarmai 1.0.0

A highly flexible multi-agent ReAct framework powered by Google Gemini

🐝 Swarm

Swarm is an advanced, open-source multi-agent intelligence framework designed for autonomous collaboration. Powered by Google's Gemini 3.1 Flash Lite, Swarm allows multiple agents to interact, use tools, execute code, and perform web searches in a shared environment.

🚀 Key Features

• N-Agent Collaboration: Run any number of agents in a single session.
• Autonomous Tool Usage: Agents can perform Google searches and execute Shell commands.
• Agent Sandboxing: Per-agent blacklists for terminal commands and file paths.
• Turn Management: Agents can autonomously pass turns to specific peers using pass_turn.
• ReAct Loop: High-level reasoning before acting, supporting multi-step tool execution.
• Files API Support: Agents can upload and analyze PDF, images, and logs via Google Files API.
• Persistent Sessions: Save and load full agent histories and environment states.

🛠 Installation

# Clone the repository
git clone https://github.com/ProgVM/swarm
cd swarm

# Install dependencies
pip install .

Or:

pip install swarmai

💻 CLI Usage

Start a basic conversation with two agents:

swarm --keys YOUR_API_KEY --agents_count 2

Start an autonomous coding session:

swarm --first_msg "Write a web scraper in Python and test it" \
      --sys1 "You are an Expert Coder" \
      --sys2 "You are a Security Auditor" \
      --cmd_timeout 600

Advanced Configuration

Swarm supports individual parameters for every agent using aiN_ prefixes:

• --ai1_name "Architect"
• --ai1_model "gemini-2.0-flash"
• --ai2_name "Tester"
• --ai2_temp 0.2

⚙️ Configuration File

You can use a JSON config to manage complex Swarms:

{
  "agents_count": 3,
  "model": "gemini-3.1-flash-lite",
  "cmd_blacklist": ["rm -rf", "format"],
  "no_pause": true
}

Run with: swarm --config my_swarm.json

⌨️ Command Center (Interactive Menu)

Press Ctrl+C at any time to:

• Rotate API Keys.
• Toggle reading pauses.
• Save the current state.
• Change logging levels (DEBUG, INFO, WARNING).
• Inject manual directives into agents' minds.

🤝 Contributing

Developed by ProgVM.

• Email: [e-mail]
• GitHub: https://github.com/ProgVM

License: MIT