kiwi

A review gate for AI-generated plans. Your coding agent writes a plan; kiwi opens it in your browser; you select text and leave comments like a Google Docs review; your feedback lands directly back in the agent's context.

Quick start

npx kiwi-review review plan.md

Your browser opens with the rendered plan. Select any text, click 💬 Add comment, write your note. Hit Send feedback — kiwi prints your review to stdout and exits. No comments? Hit Approve.

Wire it into Claude Code (recommended)

npx kiwi-review init          # this project
npx kiwi-review init --user   # every project

This installs an ExitPlanMode hook: whenever Claude Code finishes a plan, a review tab opens automatically. Comments are sent back to Claude, which revises the plan; approval lets it proceed. You never type a command per review.

Any other agent

Add one line to your agent's instructions (AGENTS.md, rules, etc.):

After writing any implementation plan, run npx kiwi-review review <plan-file> and incorporate the feedback before proceeding.

Every agent that can run a shell command can use kiwi — the feedback is the command's stdout. There's also a Copy feedback button for pasting a review into any chat by hand.

CLI

kiwi review <plan.md> [--json] [--timeout <seconds>] [--no-open]
kiwi init [--user]
kiwi hook                     # used by the installed Claude Code hook

Exit codes: 0 feedback delivered (including approvals), 1 usage or file error, 2 review abandoned (Ctrl-C or timeout).

How it works

One shot per review: kiwi review starts a loopback-only server on a random port, opens your browser, and blocks. Comments autosave server-side (closing the tab loses nothing — reopen the printed URL). Submitting resolves the session; the CLI prints the feedback and exits. Nothing runs in the background, nothing is stored, nothing leaves your machine.