Why
Why I built it
At first I just wanted to drive AI CLIs away from my desk. Using it daily, the real challenge was keeping PTY state, native sessions, queues, permissions, and schedules intact so I could take over cleanly from anywhere.
What I built
What I built
- Normalizes differences across CLIs in screen rendering, input, reply format, and completion detection while keeping each instance’s own PTY and native session.
- Discord / Telegram relay with message splitting, FIFO queuing, busy protection, timeouts, session recovery, and attachments.
- Web Control Hub with live screen view, CJK input, split layouts, manual takeover, and throttled updates for idle tabs.
- PM2 service and cron schedule control, scoped strictly to services defined in code.
Engineering
Problems I worked through
- Terminal activity after a prompt does not mean the run has finished; busy detection, input prompts, and session continuation prevent new messages from colliding with active work.
- Web, Discord, and Telegram share the same live CLI session, so I can step away from my desk and pick up the exact thread without losing context.
Role & AI
How I split the work with AI
Every quirk and snag from daily use feeds directly into the state rules and safety bounds; AI coding agents write most of the TypeScript and tests.
Trade-offs
What testing made me change
- Kept each CLI’s native stateful session instead of wrapping it as an OpenAI-compatible API—harder to integrate, but preserves full features and avoids friction with provider terms.
Outcome
Outcome
The core source stays private; I published a sanitized repo with the architecture diagram and an interactive demo.