Release of the configuration
Process 1Every connection carries an authorised group or person. Anyone outside it cannot get past the address, even if they know it.
MCP server for your ERP: teamspace in Claude, ChatGPT and Cursor.
The teamspace MCP server connects the cloud ERP teamspace with AI assistants such as Claude, ChatGPT and Cursor. An MCP server is the interface through which an AI assistant works with a piece of software under the Model Context Protocol (MCP): the server offers tools, and the assistant calls them when a question needs them. It is included in every edition. You ask in plain language, your assistant fetches the answer from your tenant, and it does so with exactly your rights.
In brief
What an MCP server does, and what it does not.
MCP stands for Model Context Protocol, an open standard through which AI assistants work with other software. Anthropic introduced it at the end of 2024 and published it openly. The MCP server tells the assistant which tools exist, for example "search tickets" or "project information". When you ask a question, the assistant decides for itself which tools to call and turns the results into an answer. Because MCP is a standard, teamspace is connected once and then works with the assistant of your choice. If you switch provider later, the connection stays the same.
In the chat it looks something like this. You ask: "Which projects have used more hours than planned this month?" The assistant answers with the current figures from teamspace. You write: "Please create an open item for Anna on that, due Friday." Shortly afterwards the open item is in teamspace, without you having opened the application. Creating it requires two things: your MCP connection includes a writing tool for open items and has been enabled for you, and you are allowed to create open items for colleagues in teamspace yourself.
Behind this lies a simple idea. A report answers the question someone anticipated when setting it up. An assistant answers the question you have right now, even across several areas: tickets, projects and contacts in one question, without opening three views.
It matters just as much what MCP is not. It is no replacement for reports that are needed regularly, and no guarantee of identical wording, because an assistant phrases things afresh every time. Nor is it a technical system account: in teamspace, an MCP connection is always tied to a person and their rights.
Examples
Questions you can put to your assistant.
The questions look quite different depending on the role. Each uses tools from the teamspace MCP server's toolkit and requires two things: an MCP connection with the right tools has been enabled for you, and you have the necessary rights in teamspace itself. What you see in the answer depends on those rights.
"Who am I in teamspace?" The quickest check that the connection works. The Who am I tool also lets the assistant understand "my tickets" or "my team".
"What is on my agenda today?" The signed-in user's agenda, right in the chat, without opening teamspace.
"Show me all unprocessed bug tickets, grouped by status." The assistant first fetches the ticket channel, then the statuses, then the tickets. Nobody had to build a report for this beforehand.
"Who is our contact at Bauer KG, and which tickets are open there?" Contact and tickets linked in one question. Without MCP, that would be two views in two modules.
"Which invoices have been unpaid for more than 30 days?" A typical question from management before the next round of reminders. The assistant looks up the documents and lists the open invoices with client and amount; asked more narrowly, just for Müller GmbH before you call them. It only works if your MCP connection includes tools for documents and you are allowed to see invoices in teamspace yourself.
"Who is on holiday next week?" The HR team's question, answered through the tools for attendance and personnel. Tools like these belong on a separate connection enabled only for HR. In any case, people only see others' absences if they may see them in teamspace.
"What have I worked on over the last four weeks?" The Activities for the last month tool supplies the material for a review or the weekly report.
In practice
The next set of minutes starts from the last one.
Weekly marketing meeting, project review, team catch-up: anyone who meets regularly has a pile of minutes after a few months. If they are kept as Markdown files in the teamspace file store, an AI assistant can work with them directly. And the next set nearly always starts from the previous one: strike what is done, carry over what is open, add what is new.
For this your administrator sets up a separate MCP connection, say "Minutes MCP", with the tools for file contents. They read, create and edit text files such as Markdown. Then you tell your assistant, for example:
- "Show me the last minutes from the marketing meeting."
- "Create new minutes for today from them. Drop items 2, 5 and 6."
- "Add a new item: sort out the stand for the autumn trade fair."
The assistant reads the old minutes, creates a new Markdown file and enters the changes. Instead of copying items across by hand, you say in one sentence what has changed since last time.
For every connection and every single tool you decide which operations are allowed: read, write, delete. The minutes connection gets read and write for file contents and nothing else, released for instance only to the people in that meeting. More on this further down under writing and audit.
Setting up
Five steps from configuration to first answer.
The administrator puts the connection together in teamspace. Users only enter an address in their client and give consent once.
- 1
Create a connection
Under Configuration, Interfaces, MCP configurations: name, language, description and the authorised group or person allowed to use it.
- 2
Choose tools
By category from the toolkit, keeping an eye on the token estimate in the connection's profile.
- 3
Share the address
Each connection has its own address following the pattern https://your-tenant/mcp/identifier. It can sit on the intranet; it is not a password.
- 4
Link the client
Enter the address as a connector, check and allow on the teamspace consent page. Client ID and secret stay empty; teamspace registers the client itself.
- 5
Ask questions
Test with "Who am I in teamspace?". If your name comes back, it works.
Several connections
One connection per purpose, not one for everything.
The teamspace MCP server is not a fixed, unchangeable offering but a toolkit. Your administrator puts together as many connections as your organisation needs. Anyone in two groups simply links both in their client. A connection belongs to exactly one tenant: anyone working in several tenants gets a separate connection for each.
-
Its own address, tools and release
A general connection for everyone with read-only tools, an HR connection for the HR department only, a sales connection with CRM tools for the sales team. Each gets its own address and its own permission group.
-
Language per connection
Each configuration has a language in which its tools are described. For an English-speaking team, you create a connection of its own.
-
Active as the kill switch
Remove the Active flag and the connection stops responding immediately, including for clients that are already linked.
-
Changes take effect at once
If the administrator adds or removes a tool, the client sees the new list the next time it connects. Nobody has to set up their connector again.
Security
Three barriers before any access. The last one cannot be switched off.
Any access through MCP has to clear three independent barriers. If one of them fails, nothing happens.
OAuth consent
Process 2On first connection, teamspace asks what the client may do. Without that consent there is no access. It applies per client and can be withdrawn again.
The user's rights
Process 3The normal permission profile applies to every single call, exactly as in the interface. Anyone who may not see a personnel file in teamspace will not see it through an AI assistant either. Two colleagues asking the same question therefore get different answers. You do not need a second permission concept.
Writing and audit
Reading, calling tools and writing are three separate permissions.
On the consent page, teamspace lists three permissions separately: read MCP server, so the client knows which tools exist, call tools, to run them, and write data. Writing is therefore a permission of its own, granted separately. The usual approach is to build the general connection with read-only tools and keep writing tools for a separate connection with a narrower release. An assistant can get things wrong, and it has no sense of whether a change had consequences.
The page also shows the redirect address, for Claude for example https://claude.ai/api/mcp/auth_callback. It is the most important checkpoint: if the domain does not match the service you are connecting, cancel.
For administrators there is the audit of OAuth permissions under Configuration, Interfaces: who allowed which client what, and when. Alongside it are the registered OAuth clients with the permissions actually granted, and the tenant client policies that set which clients are allowed in the tenant at all. Take a right away from someone in teamspace and it is gone through MCP straight away too. You never maintain permissions twice.
Short call
Which connections do your teams need?
In a first call we go through who in your organisation wants to ask what, and sketch the right MCP connections with tools and releases.
Supported clients
The AI assistants we have tested with teamspace.
Sign-in runs over OAuth, a standard MCP-capable assistants support. In principle, then, any client can be connected. The list names the clients we have tried out ourselves, and where the address goes in each.
- Claude (desktop and web)
- Settings, Connectors, Add custom connector The reference route, described in full. Desktop and web behave the same.
- ChatGPT (web)
- Settings, Connectors Use Firefox or Safari; in Chrome the sign-in currently breaks off.
- Codex (OpenAI desktop)
- MCP server configuration in the app
- Cursor
- Settings, MCP Setting up through the interface and through the configuration file both work.
- VS Code
- MCP settings Automatic discovery via the discovery settings, or enter the address by hand.
- LibreChat
- Add MCP server Self-hosted; also asks for transport and authentication. The LibreChat administrator decides which MCP servers are allowed.
- Kilo Code
- CLI or configuration file Pay attention to the OAuth permissions during setup.
Tailoring tools
Two connections with twelve tools beat one with twenty-four.
When it connects, an AI assistant receives the complete list of tools and has to keep it in mind for every question. That list takes up room in the model's working memory. The more tools, the longer each answer takes, and the more often the assistant reaches for the wrong one among similar tools. A few well-chosen tools give better answers than many, most of which have nothing to do with the question.
teamspace shows you this while you put a connection together. Each connection's profile carries a token estimate, such as "42.2k tokens (fits comfortably)" for a connection called Staff with 36 tools for everyday tasks. Tailoring works best from the task: write the use case down in one sentence, take only the tools that sentence needs, and add Who am I.
Choosing tools is not a security setting. It cuts a connection to its purpose and makes it faster. Who sees what is decided by the permission profile alone. Your administrator can also give each tool a name of its own, because the name is the text the assistant reads.
The toolkit
The areas you put tools together from.
The tools are grouped by area. For each tool you decide whether it may read, write or delete. The catalogue keeps growing; what counts is the selection in your tenant.
Context and everyday work
- Who am I
- Agenda
- Activities for the last month
- Open items and appointments
- Files and file contents: read, create and edit Markdown and text files
Projects and service
- Projects and project roles
- Project times and capacity
- Tickets, ticket channels, ticket processes and ticket statuses
- Expenses
- Documents and document line items
Clients and staff
- Contacts in the CRM
- Sales
- Search staff and staff information
- Attendance
- HR: leave request, sick note, HR data
Editions
Included in every edition, even where there is no API.
The MCP server is included in every teamspace edition at no extra charge. The REST API, by contrast, is only available from the enterprise edition. For many customers on the office edition, MCP is therefore the first way to connect teamspace to other tools at all.
The two routes solve different problems. The API is for software that always does the same thing, such as the nightly handover to a BI tool. MCP is for questions you do not know in advance. Once a question turns out to be asked the same way every week, that is the moment to build it properly as a report or via the API. The side-by-side comparison is on the AI in teamspace page; applications of your own built on the API are covered under vibe coding.
Connect and extend
What goes with the MCP server.
AI in your ERP Your AI, your data, your rights: the stance behind every AI feature. Vibe coding Build apps of your own with AI help that work with teamspace through the REST API. Integrations & API DATEV, Microsoft Teams, Outlook and the open REST API. Security Hosting in Frankfurt am Main, encryption and data protection.
First call
Watch your assistant work with teamspace.
In the call we link an MCP connection to Claude or ChatGPT and ask the questions your team actually has. You will also see the consent page and the audit.
Frequently asked questions about the MCP server
What is an MCP server?
An MCP server gives an AI assistant tools, under the Model Context Protocol, with which it accesses a piece of software. The teamspace MCP server offers tools for tickets, projects, contacts, agenda and more. The assistant calls them when you ask a question and answers with data from your tenant.
Which AI assistants work with the MCP server?
We have tested Claude (desktop and web), ChatGPT (web), Codex, Cursor, VS Code, LibreChat and Kilo Code. Because sign-in runs over OAuth, any MCP-capable client can in principle be connected. Details under supported clients.
Which edition includes the MCP server?
Every edition, at no extra charge. The REST API, by contrast, is only available from the enterprise edition. Which route suits what is set out on the AI in teamspace page.
Can someone gain more rights through an AI assistant?
No. Three barriers stand in the way: releasing the connection to a group or person, the OAuth consent, and the signed-in user's rights. The last one cannot be switched off. An MCP connection can never do more than the person using it.
So where does the data go when I ask through MCP?
To the AI assistant you have chosen yourself. On the MCP route the AI works in your assistant, not in teamspace, and that provider's terms apply to the processing there. teamspace itself runs no language model and passes no data on to other services. So you decide which provider you trust. How teamspace handles your data is set out on the security page.
Can the AI change data in teamspace through MCP?
Only if that is allowed. Writing is a separate permission on the consent page, and the administrator has to add writing tools to a connection deliberately. The usual set-up is a read-only connection for everyone and a separate connection with a narrower release for writing tools.
How do I switch an MCP connection off?
With the Active flag in the MCP configuration. A deactivated connection stops responding immediately, including for clients that are already linked. Individual users remove the connector in their client, and the administrator can also revoke the permission on the server side.
Want to connect teamspace to your AI assistant?
In a short first call we work out which teams ask which questions, and what the right MCP connections with tools and releases look like.