Connectors and MCP
Integrations: connecting Kisuno to your systems
Five connections are prepared and ready to use without any development work: your ERP, SharePoint from Microsoft 365, web search with source references, the generation of Excel, Word and PDF files, and your own documents as a knowledge base. They are activated and released during setup in the admin panel. Every further system with a programming interface is connected via the Model Context Protocol.
In brief
- Five connections are prepared and are activated and configured per customer: your ERP, SharePoint and Microsoft 365, web search with source references, file generation for Excel, Word and PDF, and your company documents as a knowledge base.
- The Model Context Protocol (MCP) is an open standard. Any system with a programming interface can be connected without tying you to a single vendor.
- For widely used tools such as HubSpot, Slack, Jira, Confluence, Notion, Linear, GitHub, Sentry, PostgreSQL, Stripe and Google Drive, ready-made MCP servers exist in the open ecosystem.
- Connections with a shared service account run through the MCP gateway. There, permissions apply per server and per individual tool, evaluated in the order user, role, group, server default.
- If the person’s identity is passed through to the target system, for example with SharePoint via the Microsoft account, the target system enforces the permissions itself and its logging applies.
- Every tool call through the gateway lands in the audit log and can be exported. The Outlook add-in is in beta and in internal use today, the Excel add-in is in progress.
Which connections are prepared
These five connections need no development work. They are activated and configured during setup in the admin panel and released per person, role or group. What each of them can do is described in the following sections.
ERP connection
The ERP connection brings customers, invoices, projects and hours from your ERP into the chat. Typical everyday questions: “How many hours are booked on project X?” or “Which invoices have been open for more than 30 days?” The answer comes from your real data, not from the model’s training knowledge.
The benefit lies less in looking things up than in chaining steps: an analysis that would otherwise mean export, sorting and formulas comes about in a single sentence and returns as a finished Excel file. Which data is visible in the process is not decided by the model but by the permissions set beforehand.
SharePoint and Microsoft 365
The SharePoint connection searches your document libraries from Microsoft 365 and answers questions from them. Contracts, minutes, quotes and policies stay where they are today. Nothing is copied into a second repository.
Sign-in uses each person’s Microsoft identity. Through the chat, everyone therefore sees exactly the libraries and folders they are allowed to open directly in SharePoint. Existing permissions remain the boundary, without you having to maintain them a second time.
Web search with source references
Web search delivers up-to-date information with a link to the source, so that every statement remains verifiable. Searching runs through a self-operated search instance (SearXNG) that is part of the stack and builds no user profiles. To turn a hit into text that has actually been read, a page-fetching service is included, which we roll out during setup.
To be honest about it: the search queries are passed on to external search engines, without your user identity. The claim “everything stays in Europe” applies to storage and model processing, not to the search query itself. Anyone who does not want external queries switches web search off in the admin panel.
Generating files: Excel, Word and PDF
An instruction such as “Create the monthly report as Excel” produces a real file to download, not a table in the chat window. This applies to analyses in Excel, to reports and letters in Word, and to PDF documents.
The code runs in a sealed-off environment per session, not on your computer; the environment is discarded after the session. An egress proxy with an allow list for network access from this environment is in progress, as is an additionally hardened isolation layer.
Company knowledge as a knowledge base
Your own documents become a searchable knowledge base: manuals, process descriptions, quote templates, minutes. Uploading happens in the chat or centrally by the admin; the answer names the source so that you can check it.
The knowledge base resides on your instance and does not leave it. New employees ask the chat instead of interrupting someone, and recurring questions about internal procedures no longer cost anyone a search through the file share.
What is MCP and why does it matter
MCP stands for Model Context Protocol and is an open standard for how an AI talks to a business application. Instead of programming a separate interface for every combination of AI tool and application, there is one protocol: the application provides an MCP server, and that server tells the AI which tools it may call.
A tool here is a single, named capability: “search records”, “read invoice”, “store file”. Precisely this decomposition is why granting permissions works at a fine-grained level at all. Without named tools there would only be the choice between full system access and none.
The open nature of the standard matters. MCP has been published and is meanwhile supported by many AI applications and vendors. An MCP server that you build for your own system is therefore not an investment in Kisuno, but in your own infrastructure: it keeps working even if you switch the platform on top of it.
For you that means: no lock-in when connecting, no point-to-point interfaces to maintain and ready-made servers from the ecosystem instead of in-house builds. The most common queries about MCP, permissions and costs are answered in the frequently asked questions.
Connecting any system with a programming interface
Any system that has a programming interface can be connected via MCP. For many widely used tools, ready-made MCP servers already exist; for everything else a REST API and a thin server in front of it are enough. Both are registered and released in the admin panel.
Ready-made MCP servers for widely used tools
The following MCP servers come from the open ecosystem around the standard, partly from the vendors themselves and partly from the community. These servers are not part of Kisuno and are not maintained by us. You register them as servers, store the credentials and release the individual tools.
| System | What is possible with it in the chat |
|---|---|
| HubSpot | Look up and add contacts, deals and notes in the CRM. |
| Slack | Search channels and messages and summarise discussions. |
| Jira | Search and create issues and query the status of a sprint. |
| Confluence | Search pages and summarise the content of a space. |
| Notion | Read pages and databases and update entries. |
| Linear | Search and create issues and check ownership. |
| GitHub | Search and assess repositories, issues and pull requests. |
| Sentry | Query error reports and assess clusters by version. |
| PostgreSQL | Query databases read-only and generate analyses from them. |
| Stripe | Look up payments, subscriptions and the status of invoices. |
| Google Drive | Search files and draw on their content for an answer. |
Your own systems via REST API
For your own industry solution, the time-tracking system or the warehouse program, a REST API is enough. In front of it comes an MCP server that translates between the protocol and your interface and thereby defines which tools exist at all. The server stays lean because the business logic remains in your system.
A practical note that determines later permission granting: cut the tools finely and unambiguously. A server with the separate tools “read invoices” and “read payroll” can be released cleanly. A single free-form tool that accepts arbitrary database queries cannot be limited sensibly and therefore belongs only in tightly restricted connections.
Connected in three steps
A new connection takes three steps in the admin panel. A newly registered server is reachable immediately, without the installation having to be rolled out again.
- Register the server. Store the address of the MCP server and the credentials in the admin panel. The tools the server offers are detected automatically.
- Define access. Decide per role, group and individual person which tools of this server can be used. Set the server default to “denied”, then only what you deliberately release applies.
- Use and verify. In the gateway’s audit log you see who called which tool on which server and when, and how it was decided.
For the prepared connections the configuration already exists. There it is straight to the question of who may use which tools.
Access stays controlled
Connections through the MCP gateway are checked on every single call. The gateway sits between the AI and the connected server and decides whether the requesting person may use this server and this tool; it writes the decision to the audit log. Where the person’s identity is instead passed through to the target system, for example with SharePoint via the Microsoft account, the target system enforces the permissions itself and its logging applies.
Permissions are evaluated in this order: individual user, then role, then group, and finally the server default. The more specific rule wins. If the check fails technically, access is denied instead of let through.
You maintain roles and groups either locally in the admin panel or take them from Entra ID. Anyone removed from a group there also loses the tool permissions attached to it.
Two ways to connect a system
There are two connection modes, and the choice determines where permissions are enforced. In the first mode the person’s identity is passed through to the target system; in the second a shared service account works behind the gateway.
| Mode | Sign-in to the target system | Who enforces the permissions |
|---|---|---|
| Identity passed through | Each person signs in to the target system once themselves. | The target system with its own user permissions. If the connection additionally runs through the gateway, its tool permissions are added. |
| Service account | One technical account for all users, stored in the gateway. | The gateway alone. Tool permissions are mandatory here, because the target system sees all requests as identical. |
Where the target system has its own user accounts, the first mode is the better choice: the permissions of the target system continue to apply unchanged. The second mode suits systems without their own user management; there the gateway is the only checkpoint, and clean tool permissions are mandatory.
How roles, groups, audit log and cost budgets interact in detail is described on the Governance page.
Add-ins for Outlook and Excel
The Outlook add-in is in beta: it runs in internal use with us today and is not yet generally available. It opens as a pane next to the open email and summarises, drafts replies, translates and extracts tasks. The Excel add-in is in progress. The goal remains to make Kisuno usable directly in the mail window and in the spreadsheet, instead of switching between application and browser.
Until then, both run through the browser: paste in emails and spreadsheet excerpts, download generated Excel and Word files. The range of functions is the same, only the route is longer.
Frequently asked questions
What is an MCP server?
An MCP server is a small program that makes an application accessible to AI systems. It speaks the Model Context Protocol, an open standard, and reports to the AI a list of named tools such as “search records” or “store file”. The AI calls these tools instead of accessing the application directly.
Can we connect our ERP without everyone seeing everything?
Yes. Permissions apply per server and per individual tool, not only per system. A tool for reading invoices can be released to accounting while the tool for payroll data remains blocked for everyone except the management. If the user identity is additionally passed through to the ERP, the existing user permissions apply there as well.
What if there is no ready-made MCP server for our system?
Then you build one in front of it, provided your system has a REST API. The server translates between the protocol and the interface and defines which tools exist. The business logic stays in your system. Because MCP is an open standard, this server also keeps working with other AI applications.
Does all data stay in Europe with web search?
No, and we say so deliberately. Storage and model processing remain within Europe, but the search queries themselves are passed on to external search engines. Your user identity is not included. Anyone who does not want that switches web search off in the admin panel.
Is every access to our systems logged?
Yes, for every connection that runs through the MCP gateway. Every tool call is recorded there in the audit log, with time, person, server, tool and the gateway’s decision, including denied calls. The log can be exported and is therefore usable for audits and documentation obligations. Where the person’s identity is passed straight through to the target system, its own logging applies in addition.
Is your system on the list?
Request access now. Tell us which systems you want to connect; that feeds into the order in which we complete connections.
Last reviewed: