Skip to main content

MCP Task Server

ยท 3 min read
Andy Reid
Andy Reid
Andy the Muppit

I have never built an MCP server or published an npm package, and definitely not one with multi-agent coordination. Until 2 days ago.

Like many developers using Cursor, I rely on AI agents to help me build faster but it can be frustrating at times. Cursor has a memory system that learns your preferences, but MCP tools cannot access it. There is no API, at least not one I could find, and all my memories are locked inside Cursor's internal database with no apparent way to download them all. Which is just fantastic. I had to go through each project's agents and ask it to show me the memories from each conversation.

Every time I started a new project, I was re-explaining the same things, such as using the King's English because I am not American with their strange spellings. The preferences existed in Cursor, but my tools could not see them.

So I built a bridge. A shared context file that both the AI and MCP tools can read and write. When Cursor creates a memory, I sync it to shared context. Now every tool I build can see my preferences.

But preferences were only half the problem. Agents also lose context between sessions. They forget previous decisions, repeat the same questions, and have no record of what was tried. So I added project scaffolding that generates a memory_bank structure for each project: architecture docs, decision logs, execution tracking, and task storage. Agents read this first and have full context before they start.

What started as a workaround became a full task management system with 27 tools for managing tasks, dependencies, priorities, and multi-agent coordination.

Two hours after the initial idea, I had my first version published on npm. Cursor and Claude made this possible, and the iteration speed kept momentum strong.

Built so far:

  • 27 MCP tools for task management and coordination
  • Shared context that bridges Cursor's memory gap
  • Multi-agent support with Planner, Worker, and Judge roles
  • Auto workspace detection so it works globally

The package is live on npm. Preferences sync globally, project context persists locally. If I can do it, you can too.

I build my own tools because I trust them far more than ones downloaded.

When I need a new feature, I add it myself. I do not wait for someone else to approve a pull request or hope the maintainer agrees with my use case. My MCP has every tool I need because I built it for how I work, not how someone else imagined I might.

The barrier to building your own tools has never been lower. If a tool does not do what you need, make one that does. I can get a bit carried away though with 14 versions to date.

If you build your own tools instead of accepting limitations, I would love to connect. What have you built recently that started as "I have never done this before"?

Read more: MCP Task Server