The desk runs on your machine. It reads your mail into a queue, proposes the next step, and runs the agent when you press Run.
You install two containers: the desk and its database. Both run on your machine. You open the desk in a browser at an address on your own network.
The desk reads your mailbox, writes the work it finds into a queue, and proposes what to do next. You approve, you edit, or you send it back.
Every run starts on your machine, under your own login. It stops when the task ends.
A run launches the agent with every permission granted. The desk starts Claude Code with --permission-mode bypassPermissions. The agent does not stop to ask before it reads a file, writes a file, deletes a file or runs a command. It reaches every file your login reaches, and nothing outside it. You accept that when you press Run. The licence terms say the same thing in full.
The desk moves no money. It sends nothing outbound over the daily cap. It acts on nothing below the confidence floor. It never approves its own work.
These four hold on every run, whatever the agent asks for.
Claude Code is Anthropic’s program. Codex is OpenAI’s. You install one under your own account, and you pay that company directly. What it costs is between you and them.
Your key gives you one year of updates and one year of answers to your questions. The price page carries the four tiers and the trial.
When your key lapses nothing switches off: you keep every release dated before the lapse, and only new releases and our answers stop.