Architecture

Three-tier MCP toolkit. mcp-server provides a pluggable tool ecosystem with virtual environments and LangChain integration. mcp-client offers auto-discovery and invocation of MCP tools running on AWS Lambda. mcpengine wraps it all in a Next.js + Monaco editor UI for live tool orchestration.

Case study

MCP Stack

A full-stack MCP toolkit experiment.

The Model Context Protocol started showing up in late 2024 as the way LLMs would talk to tools. Most early MCP implementations are bare adapters — a server here, a client there, a config file in between. MCP Stack is the experiment of "what if all three layers were designed together?"

The three layers

mcp-server — pluggable tool ecosystem

The backbone. Python-based MCP server with plugin management, virtual environments per tool, and LangChain integration baked in. Production-ready patterns: Pydantic v2, API security, Docker support.

mcp-client — AWS Lambda discovery + invocation

A Python client library that auto-discovers MCP tools running on AWS Lambda and exposes them via a CLI or async API. Built so you can deploy MCP tools as serverless functions and consume them from anywhere.

mcpengine — Monaco-powered web UI

A Next.js application with the Monaco editor (the same editor VS Code uses) embedded for live tool authoring, testing, and orchestration. Iterate on a tool in the browser, hit run, see what happens.

Status

Prototype. Three repos, ~40 commits combined, all active in the same window. Polished enough to demo, not yet a unified product. Useful as a reference for "this is what an MCP-first stack looks like."

Tech stack