Connect DeskOS to Claude, ChatGPT, Cursor or any other client that speaks MCP, then ask in plain English. Occupancy, revenue, unpaid invoices, renewals, members, leads, visitors, service requests and room bookings, live from your own system, inside your own permissions.
Your normal DeskOS login. No key to generate, nothing to paste into a config file, nothing to rotate later.
Filled chips are what this person can already see in the console. Money is greyed out here because this role does not have financial access, and the last chip is greyed out for everyone: nothing can be changed over this connection.
Occupancy is on one screen, ageing receivables on another, renewals on a third, and the question you actually have crosses all three. So it becomes a message to the ops team, and the answer arrives after the decision was already made.
Desks total, sold, on notice and free to sell, per location or across the portfolio, using the same arithmetic as the live status board. Also spaces and seat counts by floor, so "where is the room to sell 20 desks" is one question.The same number the console shows
Invoices filtered to unpaid, overdue or paid, revenue by month, and the accounts behind both. Enough to answer "who owes us the most" and "what did March actually collect" without opening a report.Only if you are allowed to see money
Subscriptions coming up for renewal inside any window you name, who is already on notice, and the active contracts and leases underneath them.Ask on a Monday, not on the 28th
Member accounts and client companies with headcount, plus the clients ranked by seats held and how long they have been with you. The answer to "who are our biggest accounts" without a spreadsheet.One directory, every location
Leads by stage, owner and source, so you can ask what came in this week, what is sitting untouched, and which source is actually converting.Straight out of the CRM
Helpdesk tickets by status and location, so the facilities question gets answered in the same place as the finance one instead of in a different tab.Open, escalated or closed
Room bookings for any date, by room and by member, which answers both "is the boardroom free on Thursday" and "are we short of rooms at this centre".Any date, any location
Visitors invited, expected and already checked in, so the person covering reception can ask rather than scroll.Defaults to today
There is no tool that creates, edits or cancels anything. An assistant connected to DeskOS cannot raise an invoice, move a booking or close a ticket, because the capability does not exist to be tricked into using.
Every tool is gated by the same permission code as the console screen it mirrors. Someone who cannot open billing in the app cannot ask for invoices here, and money fields come back blank rather than estimated.
The answer is fenced to the centres that person is assigned to, at the moment they ask. A centre manager naming a site they do not run gets nothing back, not a partial total.
There is no integration project here. One connector in the assistant, one sign in, and the questions start working.
In your assistant, add DeskOS as a custom connector using the URL from your admin console. Editors like Cursor and Claude Code take the same URL in their MCP settings.
Enter your workspace slug and your usual DeskOS login. The connection is authorised by you, as you, over OAuth. No API key is ever created or pasted.
It names the locations you will be able to read and whether money will be visible, both derived from your live role rather than promised in general terms.
Start with "which locations can you see", then ask it anything above. Each teammate connects their own assistant and gets their own level of access.
Most read-only integrations are a token with the whole database behind it. This one carries the permission model, so what an assistant can read is what the person holding it could already read.
Occupancy needs the spaces permission, invoices need billing, renewals need renewals. Checked on every call, and a tool a person lacks the code for is not even offered to their assistant.
Without financial access, amounts come back empty and the assistant is told to say so rather than estimate a figure. A blank in an answer is a permission boundary, not a gap in your data.
The centres in scope are read from the live user record at the moment of the question, so a change to what someone can see applies to their next message with nothing to invalidate.
Each call is recorded against the connection that made it, with the tool it used and when. Connections are individually revocable and show when they were last used.
The Model Context Protocol is an open standard for connecting an AI assistant to a system it is allowed to read. Claude, code editors and a growing list of other clients speak it, which is why one connector works across all of them rather than one integration each.
Anything that supports MCP connectors. That includes Claude on the web and desktop, Claude Code, Cursor and the other MCP-capable editors, and ChatGPT on the plans where custom connectors are available. The connection is the same in every case.
No. Nothing is copied out and nothing is synced. Your assistant asks a specific question, DeskOS answers that question with a capped number of rows, and the answer goes back into your own chat. Your assistant provider sees the answers you asked for, exactly as it would if you pasted them in, so pick your provider the way you would pick any vendor handling operational data.
Any staff user with a DeskOS login, and they get exactly the access their role already has. Connecting is not an escalation: a centre manager connecting their own assistant still cannot see the portfolio P&L, because they could not see it in the console either.
Each connection is a revocable credential on the user, not a shared key in a config file. Revoke it and that assistant stops working immediately. Changing or removing a permission on the user takes effect on their very next question, because nothing about their access is cached.
That is where this goes, with confirmation on every action and the same permission model behind it. It is deliberately not in the first version. Read access is useful on day one and the failure mode is a wrong answer, which you can see, rather than a wrong invoice, which you might not.
No. MCP is part of your DeskOS subscription. The thinking is done by the assistant you already pay for, and we are answering questions about your own data, so there is nothing sensible to meter.
Not this way, on purpose. Answers are capped and say when they were truncated, because a chat answer is a question answered rather than an export. When you want the full table, the reports and exports in the console are still the right tool.
Bring one centre and a question your team currently answers by hand. We will connect your assistant on the call and you can ask it yourself.