You know the work you'd do every Monday if you remembered to — the competitor scan, the dependency audit, the inbox triage?

I kept meaning to. Then I'd try the hosted agents, and they'd fail on the boring part: they don't have my stuff. No keychain, no MCP connections, none of my repos, none of the skills I'd already written. A cloud runner reads a copy of my code and guesses at the rest.

I had the same problem every week, so I built Rounds.

Rounds runs Claude Code on a schedule on your own Mac — your keychain, your repos, your MCP servers, your CLAUDE.md.

You give a round a prompt, a directory, and a cadence. Daily at 6am, Mondays, the first of the month. Rounds hands the schedule to launchd and arms a small daemon to wake the Mac a minute early, so a machine asleep at 03:00 is up when its round starts. If you'd rather it never touch sleep, skip that step — rounds still run whenever the Mac happens to be awake.

The directory is the part that matters. A round runs inside a folder you choose, so Claude reads that project's CLAUDE.md and can use the skills and slash commands you already wrote for it. Most of a good prompt ends up being /your-command — the work is already written down; the round is just what runs it on Mondays.

You also choose what a round may do: read only, or edit files, or full access with no permission prompts. Read only is the default, because the setting you get for not looking should be the one that cannot break anything.

Every round writes plain files to ~/Rounds. The menu bar tells you it finished; opening it shows the dispatch as a document, with a chat about it beside it. That chat is a separate read-only session — it can answer questions about the run, and it cannot write, execute, or send.

Every round has a ceiling, in dollars and in minutes. An unattended agent with an open-ended budget is how you find out about a runaway loop from your invoice, so there isn't one. A round that fails says so from the runner itself, because the menu bar app may not be running at 3am.

I wanted something I could trust with my own machine. So Rounds has no account, no hosted runner, and nothing to log into. It drives the Claude Code you already have — it doesn't ship its own model access or ask for an API key. Your prompts and files go to Anthropic exactly as they do when you run Claude Code yourself, and nothing else leaves your Mac — no telemetry, no third party, and your credentials never move.