Show HN: A local AI workspace with dynamic routing and Docker-sandboxed

I have been working on a personal project called Black LLAB. I built this architecture because I was extremely frustrated with constantly switching between different web interfaces for different tasks. I found myself wasting time manually guessing if a prompt needed a massive cloud model or if a fast local model could handle the request. I wanted a system that could intelligently grade the complexity of my prompts and route them automatically.

Another major issue was agent safety. I wanted to experiment with autonomous AI agents that could write code and scrape the web; however, I did not want to risk running them directly on my host machine. I needed a way to replicate the secure sandboxes used by frontier labs. I solved this by integrating a Docker environment that boots up an isolated container for every agent task. This allows the models to execute code and modify files in a completely secure space.

I combined these solutions into a single open source workspace. The system currently routes simple tasks to local models like Qwen and sends complex tasks to Claude Opus or Perplexity Sonar. I am a solo developer and this is my first time releasing an architecture of this scale. I would deeply appreciate it if senior engineers could review my code and point out my mistakes. You can view the code at github.com/isaacdear/black-llab.