Asterplant is a fork of VS Code for reading and judging what a coding agent produced.
Windows 11 today. You build it from source, because there is no packaged installer yet.
Asterplant is a fork of Microsoft's VS Code, built the way VSCodium is built. A fresh upstream checkout sits at the pinned tag 1.138.0. The fork adds 28 numbered patches and its own source folders on top.
VS Code assumes a person writes the code. Asterplant assumes an agent wrote it. All 24 added features help someone read and judge work an agent produced.
Your sessions are Claude Code or Codex running in an integrated terminal. Asterplant owns that terminal and never spends your agent subscription.
Every file an agent wrote here, grouped by kind: specs, designs, plans, reports, notes, charts, scripts and other.
A red dot marks a file nobody opened since its last write. A title bar pill counts them. The count clears itself.
An HTML or Markdown file opens rendered. The page waits until the file stops changing, so a half-written page never appears.
Closing the window does not stop the agent. The next launch attaches to the same terminals, with their output intact.
After a power loss, a view lists the conversations that were open. Recover types the resume command and leaves Enter to you.
Quote to session, tab groups by artifact kind, built-in agent rules, an AI cost ledger, rule mining, and the rest.
We scored 82 products against six dimensions in September 2026. Five of them already score solved or partial on four dimensions or more. Two dimensions had no answer anywhere in the map.
Those two are the pile of reports an agent leaves behind, and the tab bar. No product in that research manages report sprawl. None fixes a tab bar full of long, agent-generated file names. Asterplant starts at those two.
Asterplant has no built-in chat view. Sessions are Claude Code or Codex in a terminal. The fork hides seven upstream surfaces, and four of them lead to that chat view.
No model gives a verdict on your code. Jev, TypeSafe's judgment model, returns typed answers and never text. It picks a file's kind when the naming rules cannot settle it. A tag you set by hand is final.
Rule mining reads what you typed, so it stays off by default. Terminal screen snapshots stay off for the same reason.
Being plain about the limit: a classification sends the file name, the folder and the file's first 40 lines to TypeSafe. Files whose names say they hold credentials are never read. That name list is a floor and never a guarantee.
Both tiers reach the same five questions. Asterplant shows the percentage of your quota you have used. A quota meter never shows a dollar amount.
The account and the quota are being built. An account you create today holds your sign-in and nothing else yet.
Email and a password today. Google and GitHub are coming.
Asterplant runs on Windows 11. There is no packaged build and no installer, so you build it from source. The scripts check the prerequisites and fetch the pinned Node. They clone VS Code 1.138.0, apply the patches and overlay the fork's source.
macOS follows once the Windows build runs daily. The licence is not settled yet.