İsmail Özyurt/ Work

12 / 12 · Tool · 2026

Ozear

A personal issue tracker with its own MCP server: AI assistants log the work and record how the code behind it works.

  • Personal tool
  • Live
  • Built with AI coding agents
Role
Product owner, AI-assisted build
Where
Personal project
Year
2026
An AI assistant calls the MCP endpoint once with log_work; the request passes through a gateway into the database and comes back as a ticket with a work note

Illustration

How it works

Presenting? ← → walk through this page, screen by screen, and on to the next project.

Built with

Every layer of the stack, and what it does in this project.

  1. L1Interface
    • Next.js
    • React
    • TypeScript

    Projects, tickets, a board and a work journal.

  2. L2Protocol
    • Model Context Protocol
    • JSON-RPC

    22 tools for assistants; accepts protocol versions from 2024-11-05 to 2026-07-28.

  3. L3Contracts
    • Zod

    One set of input rules shared by the web app and the MCP endpoint.

  4. L4Data
    • Cloudflare D1
    • SQLite
    • Drizzle ORM

    Tickets, work notes and a technology inventory with revision history.

  5. L5Hosting
    • Vercel
    • Cloudflare Workers

    The web app and MCP endpoint on Vercel; a Worker gateway is the only path to the database.

  6. L6Workflow
    • AI coding agents

    One agent implements from a written brief; a second model audits the result against the real code and tests.

What I did

  1. Designed the product: a tracker for my own projects where a work note records not just what was done but which files do what, how the data flows, why a decision was made and how it was checked.

  2. Directed the build with AI coding agents instead of writing most of the code by hand: one agent implemented from written task briefs, a second model reviewed and re-tested the result independently, and I decided what shipped.

  3. Writes are retry-safe: request keys, revision checks and atomic batches mean a repeated call from an assistant never duplicates a record or silently overwrites one.

  4. After an audit, the default search response went from 1.24 MB to 11.7 KB and the web app’s first load from 22.6 MB to 102 KB on a synthetic set of 2,000 tickets and 6,000 notes, through SQL-side paging, summaries and indexes.

  5. A single log_work call replaced a four-to-five call sequence for the most common task, recording finished work on a ticket.

  6. Deployed it myself: database migrations, the Worker gateway and the Vercel app, behind owner-only sign-in and separate tokens per AI client.

Where to next?

Next project · 01 / 12Ceviz FlowPatient cases, follow-ups and experiments in one archive: photos, DICOM, 3D models and clinical notes on one timeline.Open it