ClawCalendar
Alice active
Casting the net for the week…
read-only
Casting the net for the week…

New event

How it works

ClawCalendar

Overview

ClawCalendar is a demo of an AI agent with real write access to a shared calendar. You ask “the claw” in plain language (find an hour Bob and I are both free on Thursday and book it) and it reads everyone's week, picks a slot, and books the meeting on every participant's calendar. If that booking double-books someone, the app opens a negotiation thread between the people it actually affects.

Why this matters
1
Fewer messages, same meeting
Booking across teams usually means a chain of “does Tuesday work?” messages. An agent that can read every participant's calendar turns that into one request, and the booking lands everywhere at once.
Operations
2
Internal apps become agent tools
The calendar is exposed as an MCP (Model Context Protocol) server, which is how agents call business systems. The same pattern works for a CRM, ticketing, or inventory. Instead of a screen someone has to click through, the system becomes something an assistant can operate directly.
Integration
3
Conflicts get negotiated, not overwritten
When two commitments collide, the agent doesn't quietly overwrite someone's day. It asks the people involved and proposes a fix. The same idea covers shift swaps, room and equipment bookings, and shared resources generally.
Coordination
Under the hood
1
The request hits a locked-down agent
A message in the chat dock goes to a dedicated agent profile that only has the calendar tools. No shell, no filesystem, no arbitrary code. Even if someone hides a prompt injection in an event title, the agent can't do anything outside the calendar.
2
The agent calls the app over MCP
The app hosts its own MCP server. The agent chains calls against it: list a week, check the other participants for a free slot, then create or move an event. It acts as whichever demo user is selected at the time.
3
Changes show up everywhere, live
Every write is pushed to all connected browsers over SignalR, so both week grids re-render without a reload and the changed event pulses coral. Manual edits and the 🎲 randomize button go over the same channel.
4
Conflict threads, but only for the people involved
After every change, the app checks whether a meeting participant is now double-booked against their own event. Solo overlaps are ignored; nobody else needs to hear about those. A real conflict opens a persistent thread per pair of stakeholders, with an opening message saying who booked what on top of what. When you reply, your own claw answers in mediator mode. It can propose a move, but it can't touch an event until you agree. The thread closes itself once the two events stop overlapping.