kaizen-agent weekly downloads statistics

One tool to build and maintain your product with AI agents. Requests in, a plan you approve, the work, an independent review, done.

The problem

You ask an AI coding agent to retry a payment webhook when the bank times out. It starts editing at once, then says "Done, tests pass." The tests do pass. They were written by the same agent, from the same assumption: that a timeout means the charge failed. Sometimes the bank said yes and the reply got lost. Now a retry charges the customer twice, and nothing in the agent's report would make you look.

None of this is a lack of intelligence. It is a missing workflow:

| What goes wrong | What kaizen does |

| The agent decides what "done" means on its own, and you find out from the diff | It writes a plan first, including what wrong would look like, and waits for your approve or change | | The author grades its own work, so its blind spot is in the code and in the check | A second agent that never saw the work reviews it, from the request, the plan and the diff only | | Findings get fixed in an endless loop, or silently dropped | Serious findings are fixed and re-checked, at most twice; the rest come to you, and what is left goes to a backlog | | Requests live in sheets, docs and social posts; the plan in one chat, the review in another, the commit somewhere else | One board, from an icebox that gathers those requests to an idea, a plan, the build, the review, the fix loop and the commit, every step on file |

Many assistants now ask for approval before they act, and some run a checker on their own draft. The difference here is who checks: a reviewer that shares the author's context shares its blind spots. Kaizen's reviewer starts cold, on purpose.

Every stage is a file in .kaizen/runs//, so a run can be resumed later, by another session or another tool (Claude Code, Codex, Antigravity), and read back when someone asks why the code is the way it is.

Who it is for

Good fit if you:

• Build and maintain a product, and want one place for it: requests gathered from where they live, then plan, build, review and commit, without switching tools. • Already use an AI coding agent, and have caught it confidently wrong more than once. • Change things where a mistake is expensive: money, data, migrations, anything a customer sees. • Want to see and steer the plan before code is written, not reverse-engineer it from a diff.

Not for you if:

• Your changes are typos and one-liners. A run costs several times one prompt; kaizen takes a short path for small work, but it is not free. • You want the agent to ship without you. Kaizen is built to stop and ask. • You do not use an AI coding agent. Kaizen runs inside one; it is not one.

Install

macOS and Linux

Already have Bun? Run bunx kaizen-agent.

You also need an AI coding agent such as Claude Code, and Git for the board's commit and push.

Windows

Windows support is still in beta. In PowerShell:

Already have Bun? Run bunx kaizen-agent. As on macOS and Linux you need an AI coding agent; Git is only needed for the board's commit and push.

Use the board

The board opens in your browser. The side menu holds the board, notes, the icebox of gathered requests and settings, with a light and dark switch. From there you can start a run, approve or revise its plan, pick which review findings to fix, then commit and push.

Right after installing, kaizen setup walks you through your project, settings, this board and the phone app (it prints a QR code to scan).

Phone app

The board is a PWA: installed from the browser, it opens full screen with the kaizen icon. It needs Tailscale on this PC and on your phone, both signed in to the same account.

Scan the QR code with your phone camera, then Add to Home screen (iOS: Share menu; Android: browser menu, Install app).

By hand: run kaizen web, then tailscale serve --bg 127.0.0.1:7420, add ..ts.net to ~/.kaizen/web-hosts, and open https://..ts.net on the phone. Details in docs/guide.md. Never use tailscale funnel — it publishes the board to the internet.

To add a launcher (Spotlight on macOS, the app menu on Linux, a Desktop shortcut on Windows), run kaizen web --shortcut once.

Development

TypeScript on Bun 1.1 or newer, zero runtime dependencies. No build step for the CLI — cli/*.ts runs as-is — and the web front end ships prebuilt.

To work on kaizen, install from your checkout so every link points at it:

Skills, agents and commands are symlinked into every AI agent it finds; re-running is safe. From a checkout the workflow updates on any run, and kaizen upgrade leaves your local edits alone — update them with git.

After editing web/notebook.ts or web/notes-lib.ts, rebuild the committed bundle:

Layout

| Path | What it is |

| cli/ | The installer and the runtime behind every kaizen command: install.ts (the CLI), state.ts (runs), web.ts (board server), dashboard.ts (terminal board), sources.ts, settings.ts, notes.ts. Tests beside them. | | skills/ | The workflow as Agent Skills folders: kaizen/ (SKILL.md, the spec-.md stage contracts, adapters.md, config.default.yml), kaizen-help/, story-deck/, story-video/. | | agents/ | The Claude Code subagents: kaizen-planner, kaizen-builder, kaizen-reviewer. | | commands/ | The /kaizen- slash commands, one file each. | | web/ | The board and notebook front end: board.html, notes.html, notebook.ts. | | bot/ | Kaizen Bot — a separate private package with its own tests and build. See its README. | | docs/ | The guide and internal notes, indexed in docs/README.md. | | design.md | Design notes for the repo. | | .kaizen/ | This repo's own runs — kaizen is built with kaizen. Gitignored: a fresh clone gets it on its first run. |

Kaizen Bot

Contributing

Pull requests; CI runs bun test on Ubuntu and Windows and it must pass. For non-trivial work, dogfood: /kaizen in this repo plans, builds and reviews it — skills/kaizen/spec.md is the contract the stages follow.

License

MIT