Why I built Zippy: tickets in, verified work out
I use Claude Code every day. The agent part works well. What broke down for me was everything around it: five terminals open, agents stepping on each other's files, and no clean way to see what each one actually did before I merged it.
So I built Zippy, a local task board that hands tickets to Claude Code agents and brings back work I can check.
The loop
- Write a ticket, or sync one from Linear.
- Zippy creates a git worktree for that ticket and starts a Claude Code agent in it over the Agent Client Protocol.
- The agent works. My own checkout stays untouched.
- I get the diff, the test output and screenshots on a review card.
- I approve, and
ghopens the PR.
Why a worktree per ticket
Parallel agents in one checkout fight over the same files. A worktree gives each ticket its own working directory on its own branch, sharing one .git. Three agents can run at once and none of them can break what I'm working on.
Why local
Zippy runs on my machine with my existing claude login. No API keys stored, no cloud sandbox, no server to host. The agents see the same repo, tools and environment I do, which is where most "works in the sandbox, fails locally" problems go away.
The part that matters: review
An agent that writes code is easy now. An agent whose output you can trust without reading every line is not. Zippy is built around that gap:
- Role pipelines. Chain agents like validator → planner → executor → reviewer, so a second agent checks the first.
- Evidence, not claims. Diffs, tests and screenshots, not a summary saying "done".
- Approval cards. Anything risky, like browser actions, waits for a human.
What I'd tell anyone building with agents
Spend less time on the prompt and more on the check. The agent is the cheap part. The part that saves you is the step where a human, or a second agent, verifies the work before it ships.
Zippy is MIT licensed. If you run Claude Code on more than one thing at a time, try it and tell me what breaks.