Zippy
Open source · dev tool · built by Aziz Khan
The problem
Handing tickets to coding agents means babysitting terminals and trusting unverified diffs.
What I built
Local task board that runs Claude Code agents on each ticket in an isolated git worktree over the Agent Client Protocol, syncs from Linear, and brings back diffs, tests and screenshots to approve.
Engineering decisions
- Each ticket gets its own git worktree and branch, so parallel agents never touch the same files or my own checkout.
- It runs locally on the existing Claude Code login: no stored API keys, no cloud sandbox, no server to host.
- Role pipelines chain agents (validator, planner, executor, reviewer) so a second agent checks the first.
- Review cards show the diff, test output and screenshots. Risky actions wait for a human to approve.
Stack
- TypeScript
- Claude Code
- ACP
- Git worktrees
Written up in detail
- Why AI Agents Fail in Production and How to Fix Them
Eight ways AI agents break once real users and real data arrive, how to recognize each one, and the fix that holds.
- Why I built Zippy: tickets in, verified work out
Running Claude Code agents on real tickets, each in its own git worktree, and why the review step matters more than the agent.
Need something like this? See AI Agent Development.
More case studies
Have an AI product to ship?
Tell me what needs to be built, fixed, integrated or deployed. You work with me directly, the engineer who reads the code and ships it.