A graph-centric framework for orchestrating multi-agent systems with Vibe Graphing, unifying natural language, code, and visual drag-and-drop development paradigms.
MASFactory is a multi-agent system orchestration framework developed by BUPT-GAMMA Lab, using graph as its core abstraction. Built on Node/Edge primitives, it provides a reusable component system including Agent, Graph (subgraph), Loop, Switch, and Human nodes, organized in a four-layer architecture (Graph Skeleton → Component → Protocol → Interaction).
Its signature capability, Vibe Graphing, transforms natural language intents into executable graph_design.json workflows through human-AI collaborative iteration. Three development paradigms—natural language-driven, declarative/imperative code, and VS Code visual drag-and-drop—share a unified runtime and can coexist in the same project.
The ContextBlock protocol unifies Memory / RAG / MCP context integration with pluggable adapters. The framework includes reproduced workflows of ChatDev, AgentVerse, CAMEL, and MetaGPT, validated on seven public benchmarks. The companion VS Code extension provides topology preview, runtime tracing, and human-AI interaction.
Quick Start#
pip install -U masfactory
Requires OPENAI_API_KEY environment variable. Optional: OPENAI_BASE_URL / BASE_URL and OPENAI_MODEL_NAME (default: gpt-4o-mini). Python >= 3.10 required.
Notable Applications#
- NowWhat: AI paper summarization service converting daily paper feeds into structured briefs
- OhNoPPT: Upload a paper to auto-generate editable .pptx presentations
- ChatDev reproduction: Multi-role software development workflow
- AgentVerse reproduction: Task-solving scenarios like PythonCalculator
- CAMEL reproduction: Role-playing conversation demos
Unconfirmed Information#
- Exact PyPI version number not directly obtained (no GitHub Release published)
- ACL 2026 Demo Track paper acceptance status pending
- Specific integration list in
masfactory/integrations/not detailed in README - Support for providers beyond OpenAI-compatible APIs (e.g., Anthropic, local models) not explicitly stated
- MCP adapter implementation depth and supported MCP Server scope not specified