A 44 commit Zig project from coder routes every session through Ghostty's libghostty vt terminal core, exposing parsed screen state as JSON so scripts and agents can read detached sessions the way a human would.
Boo, a 44-commit Zig project from the coder organization, reopens a 40-year-old idea: GNU screen's detach-and-reattach model. What makes the project news-adjacent is not the multiplexing, which screen and tmux already own. It is the second promise in the Boo README: scripts and AI agents can read the screen exactly as a human would, without parsing PTY bytes or faking a terminal.
GNU screen solved a real problem: keep a long-running session alive across disconnects. tmux and Zellij refined the model. None of those tools, however, were designed for a world where the next reader of the session is often a machine. An agent that needs to check on a long-running build, or peek inside a detached REPL, typically has two options: scrape raw terminal bytes and try to reconstruct a grid, or spawn a headless terminal emulator. Both are brittle. The first loses color, cursor, and scrollback context. The second is heavy and slow.
Boo's bet, as stated in the project's repository, is to route every session through libghostty-vt, the terminal emulation core that powers the Ghostty terminal. That means Boo's own session buffer holds parsed terminal state: SGR styles, cursor position, scrollback, and the terminal modes that screen scrapers routinely get wrong. The session is no longer a stream of bytes. It is a structured view.
The user-facing primitives follow from that foundation. Boo exposes send, peek, and wait as first-class commands, plus a --json output mode that emits the parsed screen directly. The result, according to the Boo repository, is that a script can attach, query, and detach without ever opening a TTY. wait watches a session for a pattern; peek returns the current parsed view; send injects input. None of this is new in the abstract. What is new is that the read side parses real terminal state, not bytes.
For human users, the surface is intentionally familiar. The detach keybinding is Ctrl-A d, reattachment is boo attach, and a full-screen session manager lives behind boo ui. The tagline, "Sessions that haunt your terminal," gestures at the reattach model: when you come back, the screen is exactly as you left it, including colors and cursor.
For AI developers, the practical question is whether a 44-commit project is worth wiring into an agent loop. The honest answer is: maybe, with eyes open. The repo is public, has a license, a flake.nix, an install.sh that pipes over curl for Linux and macOS with BOO_VERSION to pin and BOO_INSTALL_DIR to relocate, and CI scaffolding. The README links to a releases page for pre-built binaries. It does not yet have release tags, a roadmap, or benchmarks. The README positions Boo as screen-style, which is a narrower claim than replacing tmux's full feature set. A developer evaluating it for a production agent should treat it as infrastructure to track, not infrastructure to bet on yet.
What to watch next. Show HN reception will surface early user feedback, especially on platforms and edge cases the README does not enumerate. A first release tag, even 0.1, would convert the project from a proof of concept into something an agent framework can pin. And any benchmark comparing peek --json throughput to a PTY scraper would tell the community whether the libghostty foundation actually pays for itself in real agent workloads, or whether the win is mostly architectural.