MCP (Model Context Protocol) is an open standard published by Anthropic, which enables an AI assistant to interact with other systems in the same way every time. MCP is often likened to a USB-C port for AI: a system describes what it can do once, and from then on, any AI application that supports MCP can work with it. Claude, ChatGPT, Gemini and Microsoft Copilot all support it in some form.
A connector is the way in which an AI application interacts with another system. It is a small programme that sits between them and offers the assistant a fixed list of actions. CDESK MCP is the official connector for CDESK, created and maintained by the CDESK team.
- The assistant does not have direct access to CDESK. It will never have access to your CDESK, your data or your login details. It can only request actions from the list, one at a time, and wait for the response. Anything not on the list simply cannot be done.
- It operates within your permissions. CDESK MCP logs in using your custom CDESK account, so it sees exactly what you would see, and nothing more.
- You decide which actions it is permitted to apply. Once you’ve added a connector, your AI application will list everything the connector offers and let you configure each item. The section on ‘Limited access for the AI assistant’ describes ways to narrow this down.
You add a connector once for each AI application. In the application settings, this feature is usually referred to as ‘connectors’, ‘custom connectors’, ‘MCP servers’ or ‘integrations’ – different names for the same thing.
Tools
A tool is a single, precisely defined action that the assistant is permitted to perform – ‘find a company by name’, ‘list the requests’, ‘create a task’. Each tool has a fixed set of input data; it performs this one task and returns only what you requested. In this beta version, CDESK MCP offers 64 of them (39 read-only and 26 write-enabled), plus a tool for checking connection status.
You never have to enter the names of the tools. You simply ask using common words, and the assistant determines which tools to apply and in what order. In response to the question “Which of Alfa Ltd.’s requests are still open?”, it will usually first find the company by name, then list its open requests, and finally open those individual requests for which it needs further details. Most AI applications view each tool the moment it is run, so you can see exactly what has happened – and every step has been carried out in your name and with your permissions.
Why can we trust the answers?
In the case of factual questions, the assistant must not answer from memory. They must retrieve actual records and quote from them, whilst CDESK MCP compares each quote with the actual record before the answer reaches you. Any quote that is not actually found in the record is discarded. If something has been mentioned once, you will be told that it has been mentioned once – not that it is a trend. “I haven’t found anything on that” is a perfectly valid answer that the assistant will give you, rather than making something up that sounds plausible.
When you ask for an opinion rather than a fact, the assistant will first ask whether you want just verified facts, or also its interpretation, and will select everything that is its interpretation as such.
What the CDESK MCP is not
- It is not a replacement for CDESK. Complex jobs and anything where you want to have the full screen in front of you still belong in CDESK.
- It is not a robot that acts on its own. The assistant only acts when you ask it to, and everything it records goes through the same CDESK checks as anything else.
- It is not a data export tool. CDESK MCP does not save anything to disk; each answer is re-fetched from CDESK.
- It is not a way to bypass permissions. CDESK MCP sees exactly what your CDESK account sees.
- Everyone logs in separately. You never share a session with a colleague and you never see anyone else’s data.
- Your password is not stored, and the AI never sees it. You enter it on the login page in your browser, not in the chat, and from there it goes directly to CDESK to log you in. What remains after that is an encrypted session in CDESK, which is maintained by your AI application.
- Your conversations remain the responsibility of your AI provider. Whatever you type into your AI application is processed by that provider as usual, including data from CDESK in the answers. CDESK MCP does not change this in any way, so treat it just like any other chat.