Quick note up front. Codehamr here is my side project account, my day job is running a local LLM integration business for German mid market and public utilities. Not plugging the day job, just being transparent. I do mention the side project below, since the whole post is about the setup I built around it.

Been a Claude Code power user for the last year. Solid tool, but the session limits and unpredictable quality have been wearing me thin.

I am in the blessed situation that my day job gives me access to RTX 6000 servers, which I use over the weekends / nights for personal experimentation. Out of curiosity, and partly because I wanted a fallback plan if cloud tools become unreliable, I have been testing pi and opencode against various Qwen models for the last twelve months. Not full time, just on the side. Both are great, opencode especially is a swiss army knife. But honestly until Qwen3.6:27b dropped, none of the local options closed the gap to Claude Code for daily coding.

On my RTX 6000 with 96GB Qwen3.6:27b at Q8 runs with 128k context, no issues. But honestly the 96GB are overkill for a 30B model. A consumer RTX 5090 with 32GB at Q4_M can give you a similar coding experience. If you know what you are doing and have good prompting discipline, this is the first local setup where I do not really miss Claude.

For the agent layer I wanted something radically smaller than opencode. Out of curiosity I was wondering how far I could get building one from scratch. No plugins, no MCP, no themes bullshit. The agent handles search, dependencies, file work through bash on demand. Single Go binary. https://github.com/codehamr/codehamr

Just experimentation joy on my side, happy to share as MIT open source. Use it, fork it, ignore it, whatever fits. Every step toward local LLMs is a step away from someone else owning our coding workflows. Worth a weekend or two of tinkering.