TUI-use provides a PTY and headless xterm-based framework that lets AI agents observe terminal screen state as clean text snapshots and issue keystrokes to control REPLs, TUIs, and interactive CLIs.

Primary source https://github.com/onesuper/tui-use details a daemon architecture that spawns processes in pseudo-terminals, processes ANSI via @xterm/headless, and returns snapshots containing screen text, inverse-video highlights, window title, and fullscreen detection; this mirrors Browser-Use patterns for web UIs while targeting terminal gaps where bash-level execution fails on interactive input (https://github.com/onesuper/tui-use). OpenAI's 2024 computer-use release similarly equips models to interact with GUIs via mouse and keyboard primitives, yet omits native terminal session persistence that tui-use supplies through labeled sessions and custom dimensions (https://openai.com/index/introducing-computer-use/).

Original Show HN coverage omits error-recovery loops for desynchronized terminal states and the prompt-engineering burden of translating highlights into reliable keystroke sequences; synthesis with Anthropic's agent-computer interface evaluations shows such tooling reduces reliance on brittle one-shot scripting seen in early Auto-GPT deployments, enabling sustained sessions with vim, psql, and lazygit (https://www.anthropic.com/news/building-effective-agents). The daemon-plus-snapshot model advances the agentic pattern by treating any human-designed terminal program as a controllable environment rather than requiring dedicated APIs or ORMs.