I wanted a coding tool good enough that I'd actually use it every day. For now it's a hobby and I'm the only user. But I suspect plenty of people are hacking together something similar on their own and running into the same problems I did. So here's freecode: opencode that only talks to local models. llama.cpp, no cloud, no subscription.
The first run starts a setup wizard, and every later run checks the config before starting. The code is public on GitHub, license in the LICENSE file. Install is one command: bootstrap.sh detects your GPU, VRAM and driver, then sets up llama.cpp, a model that fits, opencode and the plugins.
What it runs on
An RTX 3080 with 10 GB of VRAM. Qwen3.6-35B-A3B does about 80–110 tok/s there with a 128k context. It fits because the MoE experts live in regular system RAM.
Why the plugins
Small models fail in repeatable ways, so most of the work is catching the same slip-ups over and over. tool-input-fix repairs a malformed tool call before it runs, and breaks the loop when the model keeps retrying the same failed call. no-call-guard catches a "result" the model made up without running anything. fake-call-scrub cleans up tool calls written as plain text. danger-guard won't let through rm -rf ~, curl | bash or git push.
After every edit the model gets errors fed back to it: from a static site check (site-lint-hook) and from Python tests (test-run-hook). There's also power-cost, which shows GPU power draw and its cost in the sidebar, and since 2026-11-04 the model's live speed in tokens per second too.
How I know it works
Unattended training runs: a set of tasks checked by a script, not by me. In an overnight run Qwen 35B passed 97% of tasks and the smaller 9B 78%. The 9B also invented results without calling any tools at all.
How I actually use it
Honestly? Day to day I still mostly use Claude. Qwen isn't perfect. It handles routine work well and it saves me when I'm cut off, with no internet or no money, which happens to me sometimes. That's how it'll stay until I get a better GPU.
The biggest lesson: the strength of a local setup isn't in the model itself. It's in the frame around it: plugins, tests, and checking things with a script instead of taking the model's word for it.
More on danger-guard: the guard that won't let the model run git push. On the overnight test: 9B vs 35B on 10 GB of VRAM: who fakes results. Both in Polish.