How automateyournetwork/mcpyats works
• A2A Agent-to-Agent adapter
• ~20 MCP tool servers
• A LangGraph-driven AI orchestration agent (mcpyats)
• A Streamlit UI frontend
• Shared volume storage
Main Components
a. Frontend
• Streamlit UI (streamlit/streamlit.py) running on :8501
b. Orchestrator
• mcpyats LangGraph Agent (mcpyats/mcpyats.py) on :2024
• A2A Adapter (a2a/a2a_adapter.py) on :10000
c. Tool Servers (“mcp_servers/*”)
• chatgpt, github, slack, netbox, google_search, etc.
• Each is a standalone Docker container exposing HTTP JSON-RPC or STDIO interfaces
d. Visual Editor
e. Shared Storage
f. External Services & APIs
• LLM providers (OpenAI, Gemini)
• Cloud APIs (Google Maps, NIST, AbuseIPDB, ServiceNow, GitHub, etc.)
Relationships & Data Flow
• User → Streamlit UI → HTTP → LangGraph Agent
• LangGraph Agent
– Discovers local tools (tools/ folder)
– Discovers MCP servers via STDIO or HTTP
– Discovers peer agents via A2A Adapter (.well-known/agent.json)
– Binds selected tools to LLM → issues JSON-RPC calls
• Agent → Tool Servers → Tool Execution → Agent
• Agent / A2A Adapter → Peer Agents (delegate tasks)
• Agent & Tool Servers → shared_output volume (read/write artifacts)
• External APIs ← Tool Servers
Architectural Patterns
• Microservices: each tool is its own container
• Orchestrator + Service Bus: LangGraph as central coordinator
• Polyglot: Python (FastAPI), Node.js/TypeScript, shell scripts, Streamlit
• Tool Discovery: dynamic via STDIO or HTTP JSON-RPC
• A2A Protocol: decentralized agent collaboration
Technologies
• Docker / Docker Compose (containers)
• FastAPI & Uvicorn (A2A adapter)
• LangGraph (state graph engine)
• pyATS (network automation)
• Streamlit (frontend)
• HTTP JSON-RPC (tools)
• InMemoryVectorStore + LLMs