Back to case studies
2025Solo

Custom Agent Core with MCP

LLM core with a plugin execution layer for a developer assistant: hot reload, tool chains, and explicit context handoff.

Plugin model
Hot-reload
Execution
Context handover
Surface
CLI + API
Execution core

Plugins can evolve without resetting the whole assistant.

This is a durable execution layer, not a one-off demo.

1Plugin lifecycle is separated from core concerns
2Hot-reload keeps iteration speed high
3Cascading tool calls use explicit context handoff
Role

AI Tooling Engineer

Stack
PythonFastAPIMCPTool callingLLM APIs
Problem

Once an assistant gets more capable, plugins become hard to evolve and orchestration gets brittle.

Solution

Built an EnhancedMCP core: FastAPI server, plugin lifecycle management, hot reload, cascading tool calls, and a CLI client with execution history.

Impact

The result is a reusable core that can support new workflows and tools without a full rewrite.

What I built
  • Implemented plugin hot reload without restarting the core process.
  • Built explicit context handoff between chained tools.
  • Separated core runtime responsibilities from plugin responsibilities to keep the platform maintainable.
What this proved
  • Architecture matters more than model count in AI tooling.
  • Developer trust comes from predictable execution behavior.
Related work

More projects

2025-2026

nnzen model catalog

Solo

A live catalog with 500+ model cards that makes model research less scattered.

PythonFastAPILLM APIs
2024

Trading Automation for an In-Game Marketplace

Solo

Automation for a constrained external marketplace with strategy logic, execution control, and logging.

PythonFastAPIREST APIs / Webhooks