🚀
MassGen Orchestrator
Task Distribution & Coordination
↓
🏗️
Agent 1
Anthropic/Claude
👨‍💻
Agent 2
Claude Code
🌟
Agent 3
Google/Gemini
🤖
Agent 4
OpenAI/GPT
⚡
Agent 5
xAI/Grok
↕ Real-time Collaboration ↕
↓
🔄
Shared Collaboration Hub
Real-time Notification & Consensus
Orchestrator Central Context Hub Snapshot Storage agent1/ agent2/ User Context Paths Read-Only Write Read: All agents | Write: Final agent only Agent 1 Permanent Temp Context Agent 2 Permanent Temp Context Final Agent Selection Agent 2 → Write Access Granted ✅ ✅ Custom Context Integration + ✅ Granular Access Control + ✅ Direct Repository Access
🔧

Tech Deep Dive: Backend Abstraction Challenges

🎭 Unified Interface Challenge Claude Code CLI Context sharing across agents Gemini API Can't mix builtin + custom tools GPT-5 API changes (reasoning, streaming) Most Backends Unable to autonomously collaborate 🎯 Unified ChatAgent Protocol ⚙️ StreamChunk Normalization 🔀 Workarounds Backend-specific 🛠️ Tool Integration MCP, Web, Code 🧠 Binary Decision Framework Intelligent agent coordination and workspace sharing
🎯

Tech Deep Dive: Binary Decision Framework Solution

Agent 1 Agent 2 Agent 3 Agent 4 Vote: 4 Writing Vote: 4 Vote: 4 ⚖️ Binary Choice: vote OR new_answer Agent 2 provides new_answer → ALL VOTES RESET After Reset: Deciding Deciding Deciding Deciding Async Results: Vote: 2 Vote: 2 Writing Deciding → If new_answer: reset again
🎯 Key Innovation: Vote Invalidation Creates Dynamic Consensus
🧬

Feature: Code-Based Tools (CodeAct)

MCP tool schemas don't cloud agent context — tools become discoverable Python code

Workspace Structure

workspace/
├── servers/     # MCP tools
├── custom_tools/ # User tools
├── skills/      # Openskills
└── utils/       # Agent scripts
Tool Discovery:
Agents explore filesystem, read docstrings, import what they need

execute_command

Central MCP tool that runs agent-written Python scripts

→ Docker execution
→ Timeout & security layers
→ Workspace confinement
# Agent runs their code
execute_command(
  "python utils/report.py"
)

Why It Matters

  • → Context reduction
  • → Transparent tool access
  • → Native Python composition
  • → Parallel async execution
Research: Apple CodeAct, Cloudflare Code Mode, Anthropic MCP
🔄

Advanced Coordination Features

Task Planning

Agents create structured task plans with automatic dependency tracking.

→ 8 MCP planning tools
→ Dependency management
→ Progress: pending → done

Orchestration Restart

Winner evaluates its own answer. If incomplete, auto-restarts with improvement instructions.

→ Self-correcting feedback loop
→ Configurable max attempts
→ Tailored guidance per retry

And Many More...

→ Broadcasting - Human or LLM answers
→ Diversity - Question paraphrasing
→ Novelty - No lazy rephrasing
→ Multi-Turn - Persistent sessions
→ WebUI - Real-time visualization
Result: Sophisticated coordination that goes beyond simple voting — agents plan, adapt, and execute without conflicts.
🚀

MassGen Evolution

🏗️

Foundation Era

v0.0.1 - v0.0.3
Core framework, basic streaming,
Claude, Gemini, GPT/o, Grok
LATEST
🚀

Rapid Evolution

v0.0.3 → v0.1.23
Models, agents, agent frameworks
MCP, browsing, coding, computer use, multimodal
Memory, skills, automation, evolution
Web UI, terminal UI, interactive sessions
55 Releases
•
Foundation→Expansion
🌟

Early Adopters & Community Growth

🏢 Industrial Institutions

  • Microsoft Azure
  • Gradient
  • Sparsity
  • Leading Trading Firm
  • AgentWeb

🎓 Academic Institutions

  • Simon Fraser University
  • UC Berkeley & Santa Cruz
  • University of Pennsylvania
  • University of Sydney
  • University of Toronto

🚀 Open Source Community

Active development
Growing contributor base
Global adoption
Research partnerships
MassGen Contributors Wall
🤝 Community Contributors - Thank you!
⚡

Get Started in 60 Seconds

# Install and run
uv pip install massgen && uv run massgen

# First run: Setup wizard guides you through:
# → API key configuration
# → Creates your default config
# → Downloads skills & containers as needed

# Future runs: uses your default config, or specify another
uv run massgen "Your question"
uv run massgen --config @examples/basic/multi/three_agents_default "..." # No Docker
uv run massgen --config @examples/tools/filesystem/code_based "..." # With Docker

# Web UI: Real-time visualization of agent coordination
uv run massgen --web

✅ Supported Models & Providers

🏢 Major Providers:
Anthropic Claude • Google Gemini • OpenAI GPT/Codex • xAI Grok • ZAI GLM
🏠 Local & Extended:
Cerebras • Fireworks • Groq • LM Studio • OpenRouter • SGLang • Together • vLLM...

🤖 Agents & Frameworks

AG2 • AgentScope • Claude Code CLI • LangGraph • NLIP • Smolagent

🛠️ Advanced Tools

Browser Automation • Code Execution • Computer Use • File Operations • MCP • Memory • Multimodal • Skills • Web Search

🎨 UI/UX Interfaces

Web UI (React + WebSocket) • Terminal UI (Textual) • CLI Streaming Display • Interactive Multi-Turn • Quickstart Wizard • Session Resumption
🚀

Join the MassGen Community

💬 Join our Discord!

Want to contribute to MassGen development or research? Contact @Yu in the Discord channel.

Discord QR

⭐ Try MassGen Today

Check out the repo and give us a star if you find it useful!

GitHub QR
1 / 24