For IT and leadership
MCP servers:
connect AI safely
An MCP server gives an AI access to a system, such as HubSpot or Microsoft 365. Together with you, we decide which connections your team gets and what the AI may do in them, and we put them into production. Reading only where needed, writing only with approval.
- HubSpotContacts, deals, notesRead, drafts for approval
- Microsoft 365Mailbox, calendar, SharePointRead only
- ERPCustom MCP serverWrite after approval
What is an MCP server?
An MCP server is a small program that makes the functions of a system available to AI applications, for example “find contact” or “create meeting”. MCP stands for Model Context Protocol. Anthropic released the open standard in November 2024, and since December 2025 it has been part of the Linux Foundation. Claude, ChatGPT, GitHub Copilot and Microsoft Copilot Studio can all use MCP servers.
- The AI application
- For example Claude or ChatGPT. It decides when to call which function.
- The MCP server
- It translates between the AI and the system. It runs locally on a computer or on a server online, at the vendor or at your company.
- What it offers
- Tools that take actions, data to read (resources), and ready-made prompts.
From experiment to standard in two years.
- November 2024
The launch
Anthropic releases MCP. Most servers run locally on your own computer, and there is no common sign-in yet.
- March 2025
A common sign-in
The new specification brings online servers with a common sign-in.
- December 2025
A neutral home
MCP moves to the Linux Foundation's Agentic AI Foundation, together with OpenAI's AGENTS.md.
- Since 2025
The vendors follow
By now, HubSpot, Salesforce, Microsoft, Google, Notion and Atlassian run their own MCP servers. The specification of July 28, 2026 makes servers easier to run online.
Which systems you can connect.
Official means the vendor runs the server itself. Community means open source from third parties, which we check before use and mostly run read-only.
| System | How it connects today | What the AI does with it |
|---|---|---|
| Microsoft 365 | Official, still in preview: Work IQ for Outlook, calendar, Teams, SharePoint and OneDrive. An admin has to consent. | Search the mailbox and calendar, find documents, prepare a reply |
| Google Workspace | Official for Gmail, Calendar and Drive, still in preview | Find emails about a customer, suggest meeting times, gather documents |
| IMAP mailboxes | Community servers with a read-only mode. We run our own. | Sort inquiries from a shared mailbox and pass them to the right colleague |
| No official server. Community servers run through WhatsApp Web on a personal number. WhatsApp's terms do not allow for this, and the account can be blocked. For companies, we build our own on the WhatsApp Business API. | Answer booking inquiries and hand them to sales | |
| HubSpot | Official, online, sign-in through OAuth | Summarize a customer's history before a meeting, find open deals, create a note as a draft |
| Salesforce and Pipedrive | Official. Salesforce from Enterprise Edition, enabled by an admin. Pipedrive on all plans. | Look up deals and contacts, log activities |
| Notion, Confluence and Jira | Official. Atlassian for the cloud version only. | Search knowledge, read tickets, create a page as a draft |
| Asana, ClickUp, monday.com and Airtable | Official, sign-in through OAuth. monday.com still in beta. | Create tasks, summarize project status |
| Zendesk | An endpoint on Zendesk's domain is live; we have not found documentation yet | Read tickets and suggest replies |
| Business Central | Official from Microsoft, online version only, read-only at first | Look up stock, orders or open items |
| Lexware Office and sevDesk | No official server. Community servers or a bridge through n8n or Make, read-only and with a dedicated user. | Look up open invoices and incoming payments |
| DATEV and weclapp | No official server found. If the system has an API, we build our own. | Look up bookings and receipts, post entries only after approval |
| Stripe and PayPal | Official. Many PayPal tools move money, so only with approval. | Look up payments, prepare invoices |
| n8n, Make and Zapier | Official. All three can be connected as MCP servers and bridge systems that have no server of their own. | Start an existing workflow from the chat |
| GitHub and databases | GitHub official. For databases such as PostgreSQL there are community servers; the old reference server is archived. | Search code, query data read-only |
How we run MCP servers safely.
A connection gives the AI permissions in your system. The attacks on it are publicly documented, and so are the countermeasures.
Hidden instructions
An email, a ticket or a tool description contains commands for the AI. In May 2025, Invariant Labs showed how a public GitHub issue can get the AI to publish data from private repositories.
What we do about it A person approves everything that changes data or leaves your company.
Unknown or fake servers
A server from a public directory can do more than it claims.
What we do about it Your IT team sets a list of allowed servers. In Claude Code, for example, IT can set this centrally, and everything else is blocked.
Permissions that are too broad
If someone sets up the connection with an admin account, the AI gets admin rights.
What we do about it A separate account per connection, permissions as tight as possible, starting with read access.
Security holes in the software
In 2025, the helper tool mcp-remote had a critical vulnerability (CVE-2025-6514, severity 9.6 out of 10).
What we do about it Pinned versions, updates as part of operations, and every call logged.
MCP, API or skill: what to use when?
We use MCP where a person works with the AI inside a system. Where a workflow runs the same way every time, we write code against the API.
MCP server
When someone looks things up, searches and creates drafts while working with the AI.
“What is still open for this customer?” in HubSpot.
Code against the API
When contracts, invoices or money depend on a fixed workflow.
Our HubSpot automation creates quotes, contracts and invoices through the API, the same way every time.
Skill
When the AI needs to know how your team works, such as templates and rules.
How your team writes proposals, as instructions the AI reads when it needs them.
Since April 2026, the Thoughtworks Technology Radar has also advised against using MCP out of habit when something simpler would do.
From the first connection to day-to-day operation.
Take stock
Which systems, which data, who may do what. We start with the workflow that costs your team the most time.
Connect
Official vendor servers first. We build our own server only where none exists, tailored to exactly that workflow.
Approve
A list of allowed servers, separate accounts with tight permissions, approval for everything that writes. Your IT team keeps control of the settings.
Run
Updates, logs and new connections are part of our ongoing work with you.
How we use MCP ourselves.
Our AI works through MCP in HubSpot, in Microsoft 365 with mailbox, calendar, Teams and SharePoint, and in the browser. On top of that, we run our own MCP servers, for example for a mailbox over IMAP and for our Company Brain.
Common questions.
How does an MCP server work technically?
The AI application asks the server which tools it offers and calls them when needed. The messages use the JSON-RPC format. A local server runs as a program on the computer, and an online server runs over HTTP, usually with OAuth sign-in.
Do we need our own MCP server?
Usually not. HubSpot, Salesforce, Microsoft 365, Google Workspace, Notion and Atlassian run servers of their own. We build our own for systems without one, such as DATEV, weclapp or software you built in-house. It offers only the functions the workflow needs.
What does an MCP server cost?
The server itself is rarely the main cost. The effort goes into setting permissions, building custom servers and running them. With us, that is covered by a fixed monthly price plus implementation days as needed. We work out what that means for your systems in the Free AI Strategy Call.
Does this only work with Claude?
No. MCP is an open standard. ChatGPT and Codex add MCP servers through plugins, and GitHub Copilot in VS Code and Microsoft Copilot Studio support MCP too. We usually roll out Claude because we built our own operation with it.
Are MCP servers GDPR-compliant?
That depends on the server and the AI model, not on the protocol. We check where the data is processed and whether your data processing agreements cover it. For our own servers, EU hosting is the default. If the models also have to run in the EU, we use Azure OpenAI with a data center in Sweden, Mistral from France, or a local model.
Which system should your AI be allowed to read first?
In 30 free minutes we go through your systems. Afterwards, you will know which connection pays off and which permissions it needs.
