LoopFlow - Stop prompting Claude Code. Design the loop that prompts it.

by
LoopFlow turns Claude Code from a tool you repeatedly prompt into a workflow that runs itself. Define the process once in YAML: One agent does the work, a separate agent verifies it, and failed reviews automatically trigger another attempt. Unlike generic agent orchestrators, LoopFlow is local, open source, and built specifically for coding loops, with hard USD budget caps, persistent memory, git worktree isolation, and resumable session logs. No cloud service or extra API keys required.

Add a comment

Replies

Best
Maker
📌
Hey Product Hunt! 👋 I built LoopFlow because I found myself repeating the same cycle with Claude Code: prompt it, inspect the output, explain what failed, and then prompt it again. I wanted a way to describe that workflow once and let it run automatically. A lot of my thinking was influenced by ideas shared by Steinberger and Cherny about "designing loops that prompt agents instead of manually prompting agents yourself." Their observations made me wonder what a reusable, open-source loop runner for Claude Code would look like, and LoopFlow is my attempt at exploring that idea. LoopFlow is a small open-source CLI that lets you define these workflows in YAML. One Claude agent performs the work, while a separate reviewer independently verifies the result. If the review fails, the feedback is automatically fed into the next iteration. The loop continues until it succeeds, reaches its attempt limit, or hits a configurable USD budget. Beyond the basic loop, LoopFlow supports persistent memory between runs, isolated Git worktrees, resumable Claude sessions, and a growing cookbook of starter loops for tasks such as test fixing, release checks, documentation synchronization, and technical debt audits. Everything runs locally through your existing Claude Code installation, so there are no additional API keys, cloud services, or background daemons required. You can install it with: `npm install -g /cli` LoopFlow is MIT licensed and intentionally kept small. I'm especially interested in feedback on the loop format, ideas for new cookbook workflows, and contributions from anyone who wants to share loops that solved real engineering problems, test it on different codebases, improve the runner and verification system, or help add support for other coding agents. I'd love to hear what kinds of workflows you'd want to automate next. GitHub: