Licensed to be used in conjunction with basebox, only.
// admin
Which connector for what
Applies to
Product: Cloud · Server · Audience: Administrator · Decision maker
"A computer once beat me at chess, but it was no match for me at kickboxing." — Emo Philips
A language model on its own can only write text. Connectors give it tools: a calculator to count with, a document generator to produce files, access to the web, and access to your own systems (wiki, document management, ticketing, CRM). This page describes every connector basebox offers, what it is used for, and how organisations usually switch it on. The rule of thumb: tools everybody needs every day are switched on for everybody; access to sensitive data lives in dedicated apps that are shared only with the people who need it.
Three switches decide whether a connector is used
| Level | Who | What it decides |
|---|---|---|
| Organisation | Administrator, under Administration → Connectors | Is this connector available in the organisation at all? |
| App | Administrator or app owner | Which connectors does this app offer? An app has exactly the connectors it is given. |
| Chat | The user | Is this connector used in this conversation right now? |
Connectors to your own systems (wiki, email, CRM …) work with the user's own login. The assistant therefore only sees what that person can already see in the source system. How this is enforced: MCP permission and authorization flow.
Four connectors are useful for everyone
Calculator: exact numbers instead of plausible ones
The calculator is one of the most important tools for business use, even though it does nothing but calculate.
A language model does one thing: it predicts the next most likely word. From billions of texts it has learned how answers usually look, and from your question it produces the most plausible one. It does not calculate.
Ask a seven-year-old "what is 2 + 2?" or "3 + 3?" and the answer comes instantly, because they have heard it a hundred times. Ask "what is 1,232,455,453 + 6,456,456 × 434,563?" and the answer is "ten thousand million gazillion", or "I don't know". A language model behaves the same way, except that its guess looks like a real number.
Give the seven-year-old a calculator and show them which buttons to press, and the answer is 2,806,969,344,181, every time. The calculator connector does this for the assistant: the model recognises that a calculation is needed, hands it to the calculator and reports the exact result. Besides arithmetic it covers scientific functions, statistics and unit conversion.
Use case: Customers use it to check monthly invoices, add up cost tables, compare offers, convert units and recalculate percentages in reports.
Typical configuration: Available in all apps and chats; the user switches it on when needed.
Document generator: from conversation to file
At the end of a piece of work you usually want something to hand over: a summary, a letter, minutes, slides, a table. The document generator turns the assistant's result into a file: Word, PowerPoint or Excel, the LibreOffice (OpenDocument) equivalents, or a print-ready PDF.
Use cases: "Create a PDF with the outcome of our conversation." "Turn these meeting notes into a letter to the supplier." "Make five slides out of this analysis." "Put this comparison into an Excel sheet."
Typical configuration: Switched on in all apps and chats.
Web search: most wanted by users, most feared by security officers
Web search lets the assistant fetch current, publicly available information: news, regulations that changed last month, a supplier's product sheet, anything newer than the model's training. Almost every user asks for it first, and security officers worry about it most, for a reason: with web search, the text of the search query leaves your installation.
basebox runs web search through Staan.ai, a European search index (Qwant + Ecosia) with processing in the EU, a data processing agreement and no training on your queries. DuckDuckGo is available as an alternative provider.
The safeguards:
- Off by default. The administrator enables it per app; the user switches it on per chat.
- It switches itself off again when the user leaves the chat or signs in again. There is no permanent setting.
- Consent with a warning. Before the first search a dialog appears: "Please do not enter any confidential or personal information." It cannot be disabled.
- No identity leaves the installation. Only the search term goes out, through basebox's own connection, without the user's identity or IP address.
- Plain text only. Results reach the model as plain text, framed as data to read, never as instructions to follow.
- One provider, nothing else. The connector can reach only its search provider; all other outbound traffic is blocked.
- Audited without content. Every search is recorded in the audit log (who, when, which provider), but not the query text.
Typical configuration: Your decision. Most organisations start with a dedicated research app whose only connector is web search, shared with a defined group. We recommend this setup whenever an app also has access to internal data: web search and an internal connector in the same app is the one combination through which a manipulated web page could try to get internal data out. Many organisations later make web search available in more apps once they know how it behaves.
More on web search:
- Web search in basebox: how it works and how risks are limited
- Web search security controls: the short version for reviewers
- Web search for users: how the toggle behaves
- Web Search MCP: Risks, Controls & Customer Policy (PDF)
- Implementing a requirement: setting up a controlled pilot
Email: the inbox, read-only
Connects the assistant to a mailbox over IMAP, with each user's own login. It lists folders, counts unread messages, searches by sender, subject, text or date, and reads messages. It cannot send, delete or move anything.
Use cases: "What did the supplier write about the delivery date last week?" "Summarise the unread messages in the projects folder." "Find the invoice Ms. Weber sent in September."
Typical configuration: Available in all apps; each user enters their mailbox login once.
Daily knowledge for everyone, sensitive data in dedicated apps
The product connectors link basebox to systems you already run. The rule of thumb:
- Switch on organisation-wide what people use daily and may read anyway: internal knowledge such as the wiki, the quality-management handbook or the intranet.
- Use dedicated apps for sensitive data (HR, finance, patient-related systems), and share those apps only with the group that needs them.
Because every connector uses the user's own login, an organisation-wide connector never shows anyone more than the source system would.
| Connector | What the assistant can do | Typical use case |
|---|---|---|
| DokuWiki | Search, read and edit wiki pages | "How do we onboard a new employee?" answered from your internal wiki |
| roXtra | Search and read documents of the roXtra quality-management system | "Which SOP applies to sterile goods handling, and what does step 4 say?" |
| SharePoint | Search and browse sites, libraries and documents, limited to what the user may see | "Find the latest version of the framework contract with supplier X" |
| Nextcloud | Files (including scanned documents via OCR), calendars, contacts, notes, Deck boards, Talk, tables | "What is on my calendar on Thursday, and where is the presentation for that meeting?" |
| Confluence | Search, read and edit pages, comments | "Summarise the architecture decisions in our project space" |
| Jira | Search, read, create and update issues; boards, sprints, service desk | "Which open tickets in the current sprint are blocked?" |
| YouTrack | Search, create, comment and link issues; knowledge base | "Draft a bug report from this conversation and file it" |
| CentralStation CRM | Look up and manage contacts, companies and deals | "When did we last speak to Müller GmbH, and what was agreed?" |
Two connectors are experimental and looking for test partners
basebox does not have access to every system its customers run. Some connectors are therefore built from the vendor's public specification and tested against reference installations, but not yet against a customer installation. We are looking for partners who use these systems and want to test and refine them with us. Experimental connectors are visible only to organisations that have beta features enabled.
| Connector | What it does | Status |
|---|---|---|
| ISiK FHIR | Look up patients, their clinical documents and treatment history through the ISiK FHIR interface, which every German hospital information system must offer | Built on the gematik ISiK specification; looking for a hospital to test with |
| Zammad | Search tickets and the knowledge base of a Zammad help desk; each user sees only the tickets they are permitted to see | Tested against Zammad 7.2; looking for a customer installation |
These connectors are coming next
The list changes often, because new connectors are being built continuously.
In development: Cloverleaf / FHIR (hospital integration engine), SAP i.s.h.med, ORBIS KIS, Microsoft Exchange, Lexware, Laurin.
Under evaluation: SwissLab LIS, Dragon Medical One, PDMS, PACS (Telemis), OnkoStar, POLAVIS, SP-Expert, Valuemation / USU, Imprivata.
For all connectors: sign-in with one click (OAuth) for Email, Jira, Confluence, YouTrack and roXtra, instead of entering tokens.
If a system you use every day is missing, contact support. Customer demand decides what is built next.