M.3.G.A.I

Inspiration

We wanted a true personal AI that doesn't just chat - it ACTS. Most AI wrappers fail on headless servers, can't see images, and can't call real-world tools. We built M.3.G.A.I to fix that. We wanted Jarvis, but open-source and runnable on a $5 VPS.

What it does

M.3.G.A.I is a multimodal autonomous agent. You send it a prompt and optional image to POST /doomsday and it autonomously decides which tools to use.

Features:

• Sees and analyzes images
• Real-time web search
• Live weather for any city
• Finds any location with lat/lon + Google -Connects to Github -Sends Email messages Maps link
• Built on MCP 2.2.0 Streamable-HTTP ## How i built it Server: FastMCP 2.2.0 with streamable-http on port 4000 Client: FastAPI + LangGraph ReAct agent on port 8000 Brain: Groq LPU (openai/gpt-oss-120b) Tools: Custom MCP tools (weather_api, find_location) + LangChain tools (vision, web_search)

Flow: User -> FastAPI -> LangGraph Agent -> MCP Client -> FastMCP Server -> Tools

Challenges i ran into

MCP 2.2.0 docs are new and confusing - had to figure out streamable-http vs SSE. Making LangGraph reliably call MCP tools took hours of debugging. Handling image uploads + base64 for vision while keeping the endpoint fast.

Accomplishments that we're proud of

Fully working MCP 2.2.0 implementation - most hacks still use old STDIO. Works perfectly on headless VPS/Docker. Sub-1 second tool calls thanks to Groq. Clean architecture where MCP server is reusable by any client.

What we learned

How to build production-ready MCP servers, LangGraph state management, tool binding, and how to chain LLM reasoning with real-world actions.

What's next for M.3.G.A.I

Add more MCP tools like GitHub, Notion, Calendar, Gmail, voice I/O, and deploy as public API for other hackers.

Built With

• fastapi
• langchain
• langgraph
• mcp
• python