Small tasks
A planner splits the request into tasks of at most a few files each, so a 16K context window holds the task, its files and the answer.
vb-harness
Get real coding work out of the model on your own machine. The harness plans a feature as small tasks, has your model implement them one at a time, proves every change with your project's own checks, and re-reviews what it built with an evidence gate.
An animated illustration of a run in the app, ending with the component it built (not a screen recording).
Built around the weak spots of local models
Section titled “Built around the weak spots of local models”Your own checks
Every change is proven by your project’s type check, lint, build and tests. A gate that fails at baseline is compared, not blocking.
Evidence-gated review
A claim becomes a finding only when a test the harness runs fails because of it. Everything else is a question, clearly marked.
Every rejection explained
Broken JSON, wrong edits and hallucinated APIs are thrown away and asked again, each with a reason code the report counts.
AstroJS is the first supported stack. Any OpenAI-compatible endpoint works: LM Studio, Ollama, llama.cpp’s server, vLLM, or a hosted API. The harness is Ruby with zero gems, and Apache-2.0.
InstallRuby 3.3, git, and a model server.
Cookbook: an Astro landing pageFrom an empty folder to a merged branch with a local 16K model, every decision explained.
How it worksIntake, planning, the task loop, integration and the final review.
Need help?Custom setups, private stacks and fine-tunes.
Built by V and Bruno LLC: Astro sites & AI tooling. Work with us →