Looklook Puts Claude Code on Your Phone: An Open-Source Remote Terminal for AI Sessions

AI coding agents frequently pause for user confirmation, which breaks unattended workflows and forces you to stay glued to the terminal. Looklook gives you a secure, encrypted mobile remote that lets you approve, answer, and monitor from anywhere, turning those waits into non-blocking interruptions.

Running large tasks with Claude Code and Codex often stops midway because the agent asks for confirmation on a command or a choice. If you're away from the desk, that session waits indefinitely. Looklook, a self-hosted open-source tool, exposes the terminal session as a web page a phone can open, giving you a full terminal with a mobile keypad, conversation cards for AI output, push notifications when the AI needs input, and the ability to paste screenshots back. It works at the terminal layer, so it supports any TUI: Claude Code, Codex, Gemini CLI, OpenCode, Kimi, Qwen Code, even htop and vim, without re-authenticating with any AI service.

Building it required solving a series of awkward problems. ttyd's bundled frontend was stuck on xterm.js 5.5, which lacks synchronized output, so the screen flickered on every AI redraw; the fix was a ported frontend on xterm.js 6 with WebGL and tmux sync. tmux filters the ZMODEM control character, breaking rz/sz, so file transfers go through an HTTP panel with 8 MiB chunks instead. For remote access, WebRTC is used, but libjuice does not support ICE restart, so dropped connections are handled by suspending the session for 60 seconds and resuming at the byte level over a fresh WebRTC connection. The client also patches a libjuice ICE role conflict bug and supports TURN over TCP for UDP-blocked networks. All client code is open source under AGPL-3.0, stores data locally, and the install package is built by GitHub Actions with SHA256 checksums.

The project is the work of one developer (imythu) plus Claude Code itself, and has released over ninety versions in its first month, with features added the same day users requested them. It demonstrates how self-hosted developer tooling can adapt to the reality that AI agents often need human approval, and that humans are not always at their desks.