justagoat

I came across this project and thought it was worth sharing. It’s called ChunkHound, a local-first, offline code search engine that lets AI assistants (and humans) explore and search large codebases — semantic search, regex, MCP protocol, etc.

After speaking with the developer, I learned something pretty wild.

The entire project — specs, architecture, implementation, even the name — was generated by an AI coding agent. No human coding, just high-level prompting with a factory-style system feeding the agent tasks and specs. The agent handled the full build end-to-end.

In a way, it even indexed itself once it was done.

* Local-first semantic + regex search (Tree-Sitter + DuckDB)

* MCP server to serve AI agents (Claude, Cursor, VSCode, etc.)

* MIT licensed

Why SLMs?

* No Cloud Costs – Runs locally, eliminating expensive API calls

* Privacy First – Data stays on-device with no third-party servers

* Low Latency – Real-time AI without cloud dependency Offline AI – Works in planes, submarines, and remote areas Customizable & Efficient – Tailored AI without large-scale infrastructure

The Challenge

Running AI Locally Deploying SLMs on resource-constrained devices is difficult. Even a small 7B model requires optimization techniques like quantization and pruning to run efficiently. Local data storage is another issue—traditional AI systems rely on the cloud, but for true Edge AI, we need a local-first approach.

The Experiment

Running AI Fully Locally We built a ChatGPT-style AI assistant that runs entirely in the browser with no network connection required.

Stack

SLM (Stories15M) running fully in-browser with wllama GoatDB for local chat history storage Zero cloud dependencies This is just the beginning. SLMs and Edge AI could enable medical AI chat applications that are HIPAA-compliant, offline assistants for embedded systems, and much more.