massgen/MassGen

pip install massgen
uv pip install massgen
uv run massgen --setup  (configures API keys, Docker, skills)

Usage

• •Interactive multi-agent mode: `uv run massgen` or `uv run massgen "your question"`
• •Multi-agent run with a config: `massgen --config @examples/basic/multi/three_agents_default "question"`
• •MCP tools are configured per-agent via YAML `mcp_servers` (type `stdio` or `streamable-http`, with `allowed_tools`/`exclude_tools` filters)
• •Serve as an OpenAI-compatible API: `massgen serve` (default 0.0.0.0:4000, POST /v1/chat/completions with SSE streaming)

Key features

• ✓Multi-agent consensus: parallel processing, intelligence sharing, voting, convergence detection
• ✓15+ model backends (OpenAI GPT-5.x, Claude, Gemini, Grok, Cerebras, Together, Groq, OpenRouter, Kimi, Qwen, Zhipu, local vLLM/SGLang/LM Studio, Claude Code, Codex, Copilot, Antigravity)
• ✓MCP integration: stdio and streamable-http servers, tool whitelist/blacklist, recursive MCP execution
• ✓Textual TUI with timeline, agent cards and vote tracking (Web UI and Rich display also available)
• ✓v0.1.97 app-layer permission engine (allow/ask/deny rules, hardline blocklist, audit ledger) plus OS-level sandbox; automation mode with status.json; massgen serve OpenAI-compatible endpoint

Best for

Developers who want coordinated multi-agent (multi-model) research, planning, and coding tasks with consensus-based quality control.

Caveats

• ⚠Early stage: README warns to expect breaking architecture changes as the system is in active development
• ⚠Agents can autonomously read/write/modify/delete files in permitted directories — README gives explicit safety warnings and recommends isolated directories and backups
• ⚠Requires API keys for hosted models (OpenAI/Anthropic/Google/xAI etc.) unless using local backends
• ⚠Codex models are noted as suboptimal for MassGen coordination prompts (optimized for shorter system messages)
• ⚠Web automation use case marked 'still in test'; requires npm install @playwright/mcp@latest