I Built an AI Agent Harness That Uses Remote MCP Tools
How I connected an AI SDK agent to my deployed MCP server and finally understood what the harness around an agent actually does.
DevAssist
An MCP server that exposes project listing, code search, and git summaries to AI clients such as Cursor.
Overview
DevAssist is a Python MCP server for AI-assisted development. It lists projects under a workspace root, searches source files while skipping noise directories, and returns a read-only git summary. It started as a local stdio server for Cursor and later gained a remote Streamable HTTP transport served with Starlette and Uvicorn, including a Render deployment.
Problem
AI coding clients are more useful when they can inspect a real workspace — projects, code, and git state — through a standard protocol instead of one-off plugins.
What I built
The same tools run locally over stdio or remotely over Streamable HTTP at POST /mcp. The HTTP app adds health and service endpoints, optional Bearer auth, and a host allowlist for deployment. Tool implementations stay in a small tools/ package so the transport can change without rewriting the capabilities.
My contribution
- Implemented list_projects, search_code, and git_summary as MCP tools.
- Added Streamable HTTP transport with Starlette and Uvicorn.
- Deployed the remote server on Render and connected Cursor plus an AI SDK client.
Technologies
- Python
- MCP
- Streamable HTTP
- Starlette
- Uvicorn
- Cursor
- Render
Related writing
I Took DevAssist Remote: Streamable HTTP, Render, and Connecting MCP Clients
Part 2 — How I moved DevAssist from a local stdio MCP server to a deployed Streamable HTTP endpoint, then wired Cursor and the AI SDK to it.
I Built My First MCP Server for Cursor (And Finally Understood MCP)
How I learned Model Context Protocol by building something useful instead of another weather API demo.