CC Launchpad

Why

I use AI coding agents and I often wanted to start, stop or check them away from my desk. Claude Code can be controlled remotely, but someone has to launch it on the right computer first. CC Launchpad does that: it runs on an always-on mini PC, behind a Cloudflare tunnel, and from a web page I pick the project and start the session.

How it works

• Each session is a Claude Code process inside a tmux session. It stops by itself after a period of inactivity and can be restarted.
• Besides Claude it supports a second agent, Pi, with a mobile chat I wrote: streamed replies, collapsible tool calls, history rebuilt after a restart.
• A catalogue of OpenRouter models with the price computed on actual usage over the last 30 days, and a cost estimate taken from the conversations. In my use about 97% of the tokens are cache reads.
• An autonomous «loop» of sessions, a button to have a repository cleaned up, git worktree management and cloning from GitHub with several accounts.

Technical choices

No inbound ports other than the tunnel, access through Basic Auth and a memory limit set in systemd. The service user may use sudo only to restart the service itself. The front end is plain HTML and JavaScript with no framework: for a personal tool that was enough. Tests are targeted Python scripts and end-to-end Playwright runs with a fake agent and a fake model provider, so they need neither network nor keys.

What I did

I wrote it on my own, about 4,100 lines of Python and 1,800 of HTML and JavaScript, in 45 commits between June and October 2026. Almost all the code is written with AI agents: I decide what is needed and check the result.

Limits

It is a tool for me, not a service. Claude Code's remote control is a research preview and allows one agent per environment. Pi sessions end if the service restarts, and are then resumed. I use it every day; the repository is private.