What they did
The poster runs Claude Code on a personal and a work account and kept losing track of which terminal was doing what, which one was waiting for input and how much of each usage window was left. They built clalt, a Go CLI that signs each account in through Claude's own login, starts sessions per account in parallel and reports usage. A local three.js page shows each running session as a person at a desk who reacts to reads, edits, failing and passing commands, and who stands up and waves when it needs an answer. Claude Code wrote most of the code while the poster chose features and reviewed. Each feature began with a short plan and design questions, each fix with a failing test, and each change went on its own branch and pull request that Claude reviewed before merge. Parallel subagents took independent pieces, and Claude drove a real browser through Playwright on fake data to check the 3D view.
How it went
After about two months the project has around 2,700 Go tests and 870 JS tests and 96 merged pull requests, and it is released as MIT open source. The browser checks caught bugs the unit tests missed, such as a subagent walking over to report to an empty chair. Before going public, Claude checked the code against Anthropic's terms and scrubbed personal data from the git history. The poster says the server uses about 20 MB of RAM.
Worth knowing
It reads transcripts locally and binds to 127.0.0.1 behind a random token. It is not a way around usage limits; every account must be your own. Commenters were wary of downloading a binary, so the poster pointed to building from source. macOS and Linux only, Windows best effort.