MassGen is a cutting-edge multi-agent system that leverages the power of collaborative AI to solve complex tasks.

Multi-agent scaling through intelligent collaboration in Grok Heavy style

MassGen is a cutting-edge multi-agent system that leverages the power of collaborative AI to solve complex tasks. It assigns a task to multiple AI agents who work in parallel, observe each other's progress, and refine their approaches to converge on the best solution to deliver a comprehensive and high-quality result. The power of this "parallel study group" approach is exemplified by advanced systems like xAI's Grok Heavy and Google DeepMind's Gemini Deep Think. This project started with the "threads of thought" and "iterative refinement" ideas presented in The Myth of Reasoning, and extends the classic "multi-agent conversation" idea in AG2. Here is a video recording of the background context introduction presented at the Berkeley Agentic AI Summit 2025.

MassGen operates through an architecture designed for seamless multi-agent collaboration:

graph TB
    O[🚀 MassGen Orchestrator<br/>📋 Task Distribution & Coordination]

    subgraph Collaborative Agents
        A1[Agent 1<br/>🏗️ Anthropic/Claude + Tools]
        A2[Agent 2<br/>🌟 Google/Gemini + Tools]
        A3[Agent 3<br/>🤖 OpenAI/GPT + Tools]
        A4[Agent 4<br/>⚡ xAI/Grok + Tools]
    end

    H[🔄 Shared Collaboration Hub<br/>📡 Real-time Notification & Consensus]

    O --> A1 & A2 & A3 & A4
    A1 & A2 & A3 & A4 <--> H

    classDef orchestrator fill:#e1f5fe,stroke:#0288d1,stroke-width:3px
    classDef agent fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    classDef hub fill:#e8f5e8,stroke:#388e3c,stroke-width:2px

    class O orchestrator
    class A1,A2,A3,A4 agent
    class H hub

The system's workflow is defined by the following key principles:

Parallel Processing - Multiple agents tackle the same task simultaneously, each leveraging their unique capabilities (different models, tools, and specialized approaches).

Real-time Collaboration - Agents continuously share their working summaries and insights through a notification system, allowing them to learn from each other's approaches and build upon collective knowledge.

Convergence Detection - The system intelligently monitors when agents have reached stability in their solutions and achieved consensus through natural collaboration rather than forced agreement.

Adaptive Coordination - Agents can restart and refine their work when they receive new insights from others, creating a dynamic and responsive problem-solving environment.

This collaborative approach ensures that the final output leverages collective intelligence from multiple AI systems, leading to more robust and well-rounded results than any single agent could achieve alone.

Core Installation:

git clone https://github.com/Leezekun/MassGen.git
cd MassGen
pip install uv
uv venv

Optional CLI Tools (for enhanced capabilities):

# Claude Code CLI - Advanced coding assistant
npm install -g @anthropic-ai/claude-code

# LM Studio - Local model inference
# For MacOS/Linux
sudo ~/.lmstudio/bin/lms bootstrap
# For Windows
cmd /c %USERPROFILE%/.lmstudio/bin/lms.exe bootstrap

Using the template file .env.example to create a .env file in the massgen directory with your API keys. Note that only the API keys of the models used by your MassGen agent team is needed.

# Copy example configuration
cp .env.example .env

Useful links to get API keys:

• Cerebras
• Claude
• Gemini
• Grok
• OpenAI
• Z AI

The system currently supports multiple model providers with advanced capabilities:

API-based Models:

• Cerebras AI: GPT-OSS-120B...
• Claude: Claude Haiku 3.5, Claude Sonnet 4, Claude Opus 4...
• Claude Code: Native Claude Code SDK with comprehensive dev tools
• Gemini: Gemini 2.5 Flash, Gemini 2.5 Pro...
• Grok: Grok-4, Grok-3, Grok-3-mini...
• OpenAI: GPT-5 series (GPT-5, GPT-5-mini, GPT-5-nano)...
• Together AI, Fireworks AI, Groq, Nebius AI Studio, OpenRouter: LLaMA, Mistral, Qwen...
• Z AI: GLM-4.5

Local Model Support (NEW in v0.0.7):

• LM Studio: Run open-weight models locally with automatic server management
  • Automatic LM Studio CLI installation
  • Auto-download and loading of models
  • Zero-cost usage reporting
  • Support for LLaMA, Mistral, Qwen and other open-weight models

More providers and local inference engines (vllm, sglang) are welcome to be added.

MassGen agents can leverage various tools to enhance their problem-solving capabilities. Both API-based and CLI-based backends support different tool capabilities.

Supported Built-in Tools by Backend:

API-based backends:

uv run python -m massgen.cli --model gemini-2.5-flash "Which AI won IMO in 2025?"
uv run python -m massgen.cli --model grok-3-mini "Which AI won IMO in 2025?"
uv run python -m massgen.cli --model gpt-5-mini "Which AI won IMO in 2025?"
uv run python -m massgen.cli --backend chatcompletion --base-url https://api.cerebras.ai/v1 --model gpt-oss-120b "How hard is IMO?"

All the models with a default backend can be found here.

Local models (NEW in v0.0.7):

# Use LM Studio with automatic model management
uv run python -m massgen.cli --config lmstudio.yaml "Explain quantum computing"

CLI-based backends:

# Claude Code - Native Claude Code SDK with comprehensive dev tools
uv run python -m massgen.cli --backend claude_code "Can I use claude-3-5-haiku for claude code?"
uv run python -m massgen.cli --backend claude_code "Debug this Python script"

--backend is required for this type of backends.

# Use configuration file
uv run python -m massgen.cli --config three_agents_default.yaml "Compare different approaches to renewable energy"

# Mixed API and CLI backends
uv run python -m massgen.cli --config claude_code_flash2.5.yaml "Complex coding task requiring multiple perspectives"

# Hybrid local and API-based models (NEW in v0.0.7)
uv run python -m massgen.cli --config two_agents_opensource_lmstudio.yaml "Analyze this algorithm's complexity"
uv run python -m massgen.cli --config gpt5nano_glm_qwen.yaml "Design a distributed system architecture"

All available quick configuration files can be found here.

MassGen supports YAML configuration files with the following structure (All available quick configuration files can be found here): MassGen supports YAML/JSON configuration files with the following structure (All available quick configuration files can be found here):

Single Agent Configuration:

Use the agent field to define a single agent with its backend and settings:

agent: 
  id: "<agent_name>"
  backend:
    type: "chatcompletion" | "claude" | "claude_code" | "gemini" | "grok" | "openai" | "zai" | "lmstudio" #Type of backend 
    model: "<model_name>" # Model name
    api_key: "<optional_key>"  # API key for backend. Uses env vars by default.
  system_message: "..."    # System Message for Single Agent

Multi-Agent Configuration:

Use the agents field to define multiple agents, each with its own backend and config:

agents:  # Multiple agents (alternative to 'agent')
  - id: "<agent1 name>"
    backend: 
      type: "chatcompletion" | "claude" | "claude_code" | "gemini" | "grok" | "openai" | "zai" | "lmstudio" #Type of backend
      model: "<model_name>" # Model name
      api_key: "<optional_key>"  # API key for backend. Uses env vars by default.
    system_message: "..."    # System Message for Single Agent
  - id: "..."
    backend:
      type: "..."
      model: "..."
      ...
    system_message: "..."

Backend Configuration:

Detailed parameters for each agent's backend can be specified using the following configuration formats:

backend:
  type: "chatcompletion"
  model: "gpt-oss-120b"  # Model name
  base_url: "https://api.cerebras.ai/v1" # Base URL for API endpoint
  api_key: "<optional_key>"          # API key for backend. Uses env vars by default.
  temperature: 0.7                   # Creativity vs consistency (0.0-1.0)
  max_tokens: 2500                   # Maximum response length

backend:
  type: "claude"
  model: "claude-sonnet-4-20250514"  # Model name
  api_key: "<optional_key>"          # API key for backend. Uses env vars by default.
  temperature: 0.7                   # Creativity vs consistency (0.0-1.0)
  max_tokens: 2500                   # Maximum response length
  enable_web_search: true            # Web search capability
  enable_code_execution: true        # Code execution capability

backend:
  type: "gemini"
  model: "gemini-2.5-flash"          # Model name
  api_key: "<optional_key>"          # API key for backend. Uses env vars by default.
  temperature: 0.7                   # Creativity vs consistency (0.0-1.0)
  max_tokens: 2500                   # Maximum response length
  enable_web_search: true            # Web search capability
  enable_code_execution: true        # Code execution capability

backend:
  type: "grok"
  model: "grok-3-mini"               # Model name
  api_key: "<optional_key>"          # API key for backend. Uses env vars by default.
  temperature: 0.7                   # Creativity vs consistency (0.0-1.0)
  max_tokens: 2500                   # Maximum response length
  enable_web_search: true            # Web search capability (uses default: mode="auto", return_citations=true)
  # OR manually specify search parameters via extra_body (conflicts with enable_web_search):
  # extra_body:
  #   search_parameters:
  #     mode: "auto"                 # Search strategy (see Grok API docs for valid values)
  #     return_citations: true       # Include search result citations 

backend:
  type: "openai"
  model: "gpt-5-mini"                # Model name
  api_key: "<optional_key>"          # API key for backend. Uses env vars by default.
  temperature: 0.7                   # Creativity vs consistency (0.0-1.0, GPT-5 series models and GPT o-series models don't support this)
  max_tokens: 2500                   # Maximum response length (GPT-5 series models and GPT o-series models don't support this)
  text: 
    verbosity: "medium"              # Response detail level (low/medium/high, only supported in GPT-5 series models)
  reasoning:                         
    effort: "medium"                 # Reasoning depth (low/medium/high, only supported in GPT-5 series models and GPT o-series models)
    summary: "auto"                  # Automatic reasoning summaries (optional)
  enable_web_search: true            # Web search capability - can be used with reasoning
  enable_code_interpreter: true      # Code interpreter capability - can be used with reasoning

backend:
  type: "claude_code"
  cwd: "claude_code_workspace"  # Working directory for file operations
  api_key: "<optional_key>"          # API key for backend. Uses env vars by default.
  
  # Claude Code specific options
  append_system_prompt: ""  # Custom system prompt to append
  max_thinking_tokens: 4096                   # Maximum thinking tokens
  
  # Tool configuration (Claude Code's native tools)
  allowed_tools:
    - "Read"           # Read files from filesystem
    - "Write"          # Write files to filesystem  
    - "Edit"           # Edit existing files
    - "MultiEdit"      # Multiple edits in one operation
    - "Bash"           # Execute shell commands
    - "Grep"           # Search within files
    - "Glob"           # Find files by pattern
    - "LS"             # List directory contents
    - "WebSearch"      # Search the web
    - "WebFetch"       # Fetch web content
    - "TodoWrite"      # Task management
    - "NotebookEdit"   # Jupyter notebook editing
    # MCP tools (if available)
    - "mcp__ide__getDiagnostics"
    - "mcp__ide__executeCode"

backend:
  type: "zai"
  model: "glm-4.5"  # Model name
  base_url: "https://api.z.ai/api/paas/v4/" # Base URL for API endpoint
  api_key: "<optional_key>"          # API key for backend. Uses env vars by default.
  temperature: 0.7                   # Creativity vs consistency (0.0-1.0)
  top_p: 0.7                    # Nucleus sampling cutoff; keeps smallest set of tokens with cumulative probability ≥ top_p

backend:
  type: "lmstudio"
  model: "qwen2.5-7b-instruct"       # Model to load in LM Studio
  temperature: 0.7                   # Creativity vs consistency (0.0-1.0)
  max_tokens: 2000                   # Maximum response length

UI Configuration:

Configure how MassGen displays information and handles logging during execution:

ui:
  display_type: "rich_terminal" | "terminal" | "simple"  # Display format for agent interactions
  logging_enabled: true | false                          # Enable/disable real-time logging 

• display_type: Controls the visual presentation of agent interactions
  • "rich_terminal": Full-featured display with multi-region layout, live status updates, and colored output
  • "terminal": Standard terminal display with basic formatting and sequential output
  • "simple": Plain text output without any formatting or special display features
• logging_enabled: When true, saves detailed timestamp, agent outputs and system status

Time Control Configuration:

Configure timeout settings to control how long MassGen's orchestrator can run:

timeout_settings:
  orchestrator_timeout_seconds: 30   # Maximum time for orchestration

• orchestrator_timeout_seconds: Sets the maximum time allowed for the orchestration phase

MassGen supports an interactive mode where you can have ongoing conversations with the system:

# Start interactive mode with a single agent
uv run python -m massgen.cli --model gpt-5-mini

# Start interactive mode with configuration file
uv run python -m massgen.cli --config three_agents_default.yaml

Interactive Mode Features:

• Multi-turn conversations: Multiple agents collaborate to chat with you in an ongoing conversation
• Real-time feedback: Displays real-time agent and system status
• Clear conversation history: Type /clear to reset the conversation and start fresh
• Easy exit: Type /quit, /exit, /q, or press Ctrl+C to stop

Watch the recorded demo:

The system provides multiple ways to view and analyze results:

• Live Collaboration View: See agents working in parallel through a multi-region terminal display
• Status Updates: Real-time phase transitions, voting progress, and consensus building
• Streaming Output: Watch agents' reasoning and responses as they develop

Watch an example here:

All sessions are automatically logged with detailed information. The file can be viewed throught the interaction with UI.

agent_outputs/
  ├── agent_1.txt       # The full logs by agent 1
  ├── agent_2.txt       # The full logs by agent 2
  ├── agent_3.txt       # The full logs by agent 3
  ├── system_status.txt # The full logs of system status

Here are a few examples of how you can use MassGen for different tasks:

To see how MassGen works in practice, check out these detailed case studies based on real session logs:

• MassGen Case Studies

# Ask a question about a complex topic
uv run python -m massgen.cli --config massgen/configs/gemini_4o_claude.yaml "what's best to do in Stockholm in October 2025"

uv run python -m massgen.cli --config massgen/configs/gemini_4o_claude.yaml "give me all the talks on agent frameworks in Berkeley Agentic AI Summit 2025, note, the sources must include the word Berkeley, don't include talks from any other agentic AI summits"

# Generate a short story
uv run python -m massgen.cli --config massgen/configs/gemini_4o_claude.yaml "Write a short story about a robot who discovers music."

uv run python -m massgen.cli --config massgen/configs/gemini_4o_claude.yaml "How much does it cost to run HLE benchmark with Grok-4"

# Single agent with comprehensive development tools
uv run python -m massgen.cli --config massgen/configs/claude_code_single.yaml "Create a Flask web app with user authentication and database integration"

# Multi-agent development team collaboration  
uv run python -m massgen.cli --config massgen/configs/claude_code_flash2.5_gptoss.yaml "Debug and optimize this React application, then write comprehensive tests"

# Quick coding task with claude_code backend
uv run python -m massgen.cli --backend claude_code "Refactor this Python code to use async/await and add error handling"

MassGen is currently in its foundational stage, with a focus on parallel, asynchronous multi-agent collaboration and orchestration. Our roadmap is centered on transforming this foundation into a highly robust, intelligent, and user-friendly system, while enabling frontier research and exploration. An earlier version of MassGen can be found here.

⚠️ Early Stage Notice: As MassGen is in active development, please expect upcoming breaking architecture changes as we continue to refine and improve the system.

✅ Timeout Management System: Timeout capabilities for better control and time management

• Orchestrator-level timeout with graceful fallback
• Enhanced error messages and warnings for timeout scenarios

✅ Enhanced Display Features: Improved visual feedback and user experience

• Optimized message display formatting and synchronization
• Better handling of concurrent agent outputs

✅ Foundation Architecture: Complete multi-agent orchestration system with async streaming, builtin tools (code execution, web search), and multi-backend support

✅ GPT-5 Series Integration: Support for OpenAI's GPT-5, GPT-5-mini, GPT-5-nano with advanced reasoning parameters and verbosity control

✅ Claude Code Integration: Native Claude Code backend with streaming capabilities, tool support, and stateful conversation management

✅ GLM-4.5 Model Support: Integration with ZhipuAI's GLM-4.5 model family with enhanced reasoning display and coordination UI

✅ Local Model Support: Complete LM Studio integration for running open-weight models locally with automatic server management and zero-cost usage

✅ Extended Provider Ecosystem: Support for 15+ providers including Cerebras AI, Together AI, Fireworks AI, Groq, Nebius AI Studio, and OpenRouter

• Model Context Protocol (MCP) Integration: Enabling Claude Code and Claude Backend to support MCP for enhanced tool capabilities (v0.0.9)
• Advanced Agent Collaboration: Exploring improved communication patterns and consensus-building protocols to improve agent synergy
• Expanded Model, Tool & Agent Integration: Adding & enhancing support for more models/tools/agents, including a wider range of tools like MCP Servers, and coding agents
• Improved Performance & Scalability: Optimizing the streaming and logging mechanisms for better performance and resource management
• Enhanced Developer Experience: Introducing a more modular agent design and a comprehensive benchmarking framework for easier extension and evaluation
• Web Interface: Developing a web-based UI for better visualization and interaction with the agent ecosystem

We welcome community contributions to achieve these goals.

Version 0.0.9 focuses on Model Context Protocol (MCP) Integration, enabling Claude Code and Claude Backend to support MCP for enhanced tool capabilities and interoperability. Key enhancements include:

• MCP Support for Claude Code (Required): 🔧 Enable MCP protocol support in Claude Code backend
• MCP Support for Claude Backend (Required): 🔧 Enable MCP protocol support in Claude Backend
• MCP Tool Integration (Required): 🔗 Seamless integration with MCP-compatible tools and servers
• Advanced MCP Features (Optional): Tool management, authentication, and orchestration patterns

For detailed milestones and technical specifications, see the full v0.0.9 roadmap.

We welcome contributions! Please see our Contributing Guidelines for details.

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.