The Ghost in the Machine: Unpacking wandb-hackathon-demo

How a local-first issue tracker built for a Tokyo hackathon reveals the strange, highly structured future of agent-generated codebases.

6 min read • View on GitHub • More from hironow

A mechanical robotic hand meticulously carving a miniature replica of a complex factory building inside a small sealed glass bottle. — Agentic development creates localized, contained versions of complex enterprise systems.

A local MCP (Model Context Protocol) server for issue tracking, built with Bun + Drizzle ORM + SQLite.

hironow, Author/Maintainer · wandb-hackathon-demo/apps/local-issues

Key Takeaways

The repository operates as a synthetic codebase where autonomous agents leave behind a forensic trail of markdown journals instead of traditional source control commits.
It provides a local-first Model Context Protocol (MCP) server that mimics cloud issue trackers, allowing AI assistants to operate without leaving localhost.
The architecture relies on Bun and SQLite to guarantee the zero-latency execution necessary for uninterrupted multi-step LLM tool calling.
This project signals a shift toward machine-to-machine infrastructure where AI agents construct and consume their own headless tooling.

An Archaeological Dig in a Synthetic Codebase

Opening this repository feels like an archaeological dig into an alien civilization. Instead of a typical human git history, the root directory is cluttered with `.expedition`, `.gate`, and `.siren` folders. These contain over 80 machine-generated journal entries and routing metadata. This is not a standard application; it is a synthetic codebase. It reveals exactly how autonomous coding agents built a local-first application.

A close up of a heavy brass magnifying glass focused on an open ledger book where the entries are mechanical punch card patterns. — The repository metadata acts as a granular, step-by-step log of the project's non-human evolution.

The Local-First Illusion

The agents constructed a local-first Model Context Protocol (MCP) server. Residing in the `apps/local-issues` directory, this server perfectly mimics a cloud issue tracker like Linear. It provides a highly structured schema to AI assistants like Claude Desktop, completely eliminating the need for API keys, OAuth, or cloud storage.

The dual lifecycles of a synthetic repository: build-time orchestration versus run-time MCP execution.

Engineered for Zero Latency

Traditional Node.js or Python servers were bypassed in favor of Bun and a local SQLite database managed by Drizzle ORM. This stack matters because LLM tool calling requires absolute minimum latency to maintain flow state. When an agent takes multiple steps to resolve an issue, network latency compounds.

Feature	Cloud SaaS API	Local MCP Tracker
Network Latency	200-500ms per call	<5ms per call
Authentication	OAuth2 / Tokens	None (Local process bound)
Storage	Remote Cloud DB	Local SQLite File
Data Privacy	Vendor access	100% Air-gapped

The Machine-to-Machine Economy

If an AI agent only interacts with software via JSON schemas, the visual UI of modern SaaS becomes dead weight. This project proves we can bypass expensive cloud subscriptions for agentic workflows. Local, headless MCP servers represent the true future of autonomous agent tooling.