Cokpit

Agent coding sessions on one PC and a home server were scattered across terminals with no shared view and no way to keep them alive across restarts.

Period
2026-06 to present
Status
Built, in use
Stack
Rust, Tauri v2, React 19, TypeScript, Vite, pnpm, tokio, axum, rusqlite
Ownership
Sole builder

Problem

Working with AI coding agents all day means many long-running terminal sessions, local and remote. Stock terminals lose them on restart, show no diff of what an agent changed, and give no single place to see every session at once. A session that lives only in one window dies with it.

Role

Kavish set the product direction, wrote the specs, reviewed every change and runs the result on his own machines. Implementation was drafted by AI coding agents from his specs and plans, under his review and tests.

Built

Sessions that outlive the app

A standalone PTY daemon owns the terminal sessions, so the window can close and reopen without losing a running agent.

A hub on a home server

Cokpit’s hub moves ownership of every agent and shell session to a daemon on a home Proxmox VM, so any client can attach, detach and come back. A headless terminal emulator keeps each session’s screen state, so a client attaching later sees a clean screen rather than a replayed stream. Each session runs in its own cgroup v2 slice, with SQLite persistence and auto-resume across daemon restarts. WebSocket and REST on loopback with bearer auth let several clients watch one session at once.

A canvas of live terminals

An infinite pan and zoom canvas shows every session at once, next to a file explorer, a Monaco editor and a live diff of what the agent changed.

A control plane for agents

An internal MCP server lets an agent in one pane list, read and drive the other panes, with every action reported back.

Result

Installed on Kavish’s workstation and used daily. Version 0.0.38 installed in October 2026. The hub daemon and client were built in one week in August 2026, with 197 daemon and 318 client tests, and run on a Proxmox VM at home; the earlier thin client was retired in favour of Cokpit as the client. Public site at cokpit.dev, private alpha.

Collaborators

AI coding agents (Claude Code, Codex): drafted implementation from specs and plans he wrote, under his review and tests. No other contributors.

What stays private

Licence keys, the licence server, the hub’s addresses and tokens, and the VM’s configuration.

Links