What they did
Jamie Duncan builds BrandCast, a large monorepo, and moved his Claude Code setup off his laptop onto an Intel NUC with 32GB of RAM. He wrote a small web app called Factory that pulls every open GitHub issue and sends it to a local model on a separate home inference box, which scores each one for complexity, impact and feasibility and writes a short rationale. Factory shows the top 15. He can run one by hand, or start auto mode: Factory starts a background Claude Code session in a fresh git worktree with his skills, MCP servers and prompts, the agent plans, edits, tests and opens a PR, and Factory polls GitHub Actions. Once the PR passes and merges, it removes the worktree, cleans up processes and branches, logs the run to SQLite and starts the next issue.
How it went
Duncan says the loop runs on its own and keeps his laptop free for research and design work. The one figure he gives is a single issue that took 151 minutes from start to merge. Scoring the backlog costs nothing in API tokens because it runs locally. He gives no count of issues closed, failure rate or token spend on the Claude side beyond being on a Max plan.
Worth knowing
He says it only works because his own framework forces tests and checks before any commit; without those, an unattended loop just piles up bad code. Every run gets its own worktree so a failed one can be thrown away. The local model only ranks issues; all code changes go through the frontier model.