Planning your day by talking to Claude, locally
There is an obvious appeal to describing your day out loud and having something else do the arranging. “Move everything after two to tomorrow, and put an hour of deep work at nine” is much faster to say than to do.
Until recently that meant handing your schedule to somebody’s server. The Model Context Protocol changes the shape of that, and it is worth understanding why.
What MCP actually is
MCP is a standard way for an AI assistant to talk to a tool. The tool runs a small server that describes what it can do — read the schedule, create a block, move one — and the assistant calls those functions when it needs to.
The important detail for a planner: that server can run on your own machine. It is an ordinary local process talking to an app already running on your Mac. There is no hosted copy of your calendar involved.
What crosses the network is what always crosses when you use a cloud assistant: your prompt, and whatever the assistant reads in order to answer it. That is not nothing, and it is worth being clear-eyed about. But it is very different from a service that continuously holds a copy of your schedule.
What it is good for
Bulk rearrangement. “Push everything after three to tomorrow morning” is one sentence and about a dozen drags.
Planning from a description. “Tomorrow: two hours of writing in the morning, a gym block at six, and keep the afternoon clear.” Faster spoken than built.
Questions about your own time. “How many hours did I actually spend in meetings last week?” Reading an analytics screen answers it; asking is quicker.
Recovering a wrecked week. “Reschedule everything I missed on Tuesday into Thursday and Friday” is genuinely tedious by hand, and exactly the sort of task people avoid — which is how a missed day becomes a missed week.
What it is not good for
Deciding what matters. It will happily reschedule the thing you have been avoiding for three weeks, forever, without ever pointing out that you are avoiding it.
Precise placement. Dragging a block is faster and more certain than describing where it should go.
Anything you would not check. It is worth reading what it did. Instructions get misread, and a schedule that quietly went wrong is worse than one you built slowly.
Setting it up
The general shape, whichever planner you use:
- Turn on the app’s local API or MCP server.
- Connect it to your assistant. Increasingly this is one click in settings — the app writes the entry into the assistant’s config for you.
- Restart the assistant so it picks up the new server.
- Ask it something read-only first, like what is on tomorrow, and check the answer before letting it change anything.
Focusly ships an MCP server inside the app and connects to Claude Desktop, Claude Code, Codex and Cursor from a button in Settings. The server is a local process reading and writing the copy of Focusly already running on your Mac.
One honest limitation worth knowing: consumer ChatGPT currently only accepts MCP servers reachable over the internet, so a purely local server cannot be connected to it without exposing it publicly — which rather defeats the purpose.
Is it worth it?
If you already use an AI assistant daily, it removes a genuine category of tedium — the bulk edits and the recovery work that you otherwise put off.
If you do not, this is not a reason to start. Dragging a block takes two seconds and it is not the part of planning that is hard. The hard part is deciding what deserves the time, and no amount of tooling has made that easier yet.