@jakevin7: Here's a hot take: TUI will gradually decline or even become obsolete. I haven't used Claude Code in a long time. I almost exclusively use Slock. For ad-hoc tasks, I now use Codex Desktop more, occasionally Claude Desktop. It makes me rethink TUI…

Here's a hot take: TUI will gradually decline or even become obsolete. I haven't used Claude Code in a long time. I almost exclusively use Slock. For ad-hoc tasks, I now use Codex Desktop more, occasionally Claude Desktop. It makes me rethink TUI as a concept. Today, the Slock group happened to be discussing TUI, and everyone's assessment of TUI was consistent: the direction is wrong. Someone said "TUI is horribly wrong," and another said "all TUIs have been led astray by Claude Code." That's a strong statement, but upon reflection, it makes sense. @OnlyXuanwo xuanwogg analyzed in the group why TUI became popular, attributing it mainly to historical reasons. Early models couldn't handle GUI well, TUI was simple to implement and could be generated by the model, so Claude Code started there. Then everyone followed the trend, and for a while, TUI became the "standard posture" for AI coding tools. But this was just path dependence, not correctness. History moves in spirals—when Claude Code started, nobody knew how interactions should work, so TUI as a starting point had its rationale. But now that models are stronger, sticking with TUI is indefensible. TUI isn't without merits—you can SSH in from anywhere, and the local app architecture is cleaner. But these advantages simply don't hold up in AI agent scenarios. For long-running tasks, complex contexts, and processes requiring visual display, the TUI experience is genuinely poor. Essentially, it's a "crippled GUI"—you have a GUI available but deliberately degrade to TUI instead. So what's the right direction? The consensus in the discussion was two paths: First, CLI + server architecture: the command line acts as a trigger, while the actual logic and state run on the server, and the frontend can be any interface. This retains the flexibility of CLI without being dragged down by the TUI experience. Second, go directly to Web UI: models can now handle it completely, there's no reason to stay with TUI. For me, after switching to Codex Desktop, task status is clear at a glance, and I no longer have to wrestle with a terminal interface. This is what an AI tool should look like. Model capabilities are evolving, and the interaction methods of tools should evolve accordingly. Still using TUI means using today's models to do what only previous models could.