kevin101681/claude-code-dashboard

Claude Dashboard is a self hosted control panel that lets you run and monitor Claude Code and OpenAI Codex, two AI coding assistants that normally live in a terminal, from any device with a web browser, including a phone. The actual work happens on your own Windows PC, which keeps running the AI agent sessions even after you close the browser tab, so you can start a task from your desktop, check its progress from your phone, and approve a permission prompt from anywhere else, then pick the same session back up later with its full history intact. There is no cloud service sitting in between and no per use API charges, since the dashboard runs the Claude Code and Codex command line tools you already have installed, using your existing subscriptions to those services. You can view a session either as a simplified chat style conversation or switch to the real underlying terminal when you need to see a raw permission prompt or a text based menu. Push notifications can alert you when a session finishes or needs your input, but only when you are not already looking at the dashboard, and tapping the notification opens that exact session. Each session also has its own permission setting you can change on the fly, so one task can run with fewer restrictions while another stays locked down. Because these AI agents can read files, run commands, and reach the network on your behalf, sometimes without asking first depending on the permission mode chosen, the README treats access to this dashboard as equivalent to a key to your whole computer. It recommends keeping it reachable only over a private network such as Tailscale rather than exposing its port to the internet, and setting a long random access token, or using a sign in service called Clerk if exposed more broadly. It also warns that an AI agent can be influenced by untrusted text it reads, such as a poisoned file or web page, so approvals should stay on for anything not fully trusted. Setting it up locally on Windows is a short sequence of installing dependencies, copying an example configuration file, setting an access token, and starting the server. The project has no build step, no framework, and no database, and the author describes it as small enough to read through in an afternoon.