Security & data
The MCP server gives an assistant the same access as the signed-in user, never more, and adds safeguards specific to an agent: read-only by default, writes through the Oxygenius API, two-step confirmations, rate limits and an audit log.
Company isolation
- The company comes only from the user's token: no tool accepts a company, database or connection string as a parameter.
- The user → company resolution is done again on every request: a deactivated account, or a company that withdraws its consent, loses access immediately.
- Each company has its own database; document files are only read under the company's own storage location, whatever path is recorded.
What the assistant sees
- The data the user sees in Oxygenius: same permissions as the application screens (details per tool).
- Other users' private events are masked; e-mails are limited to the mailboxes shared with the user when the company requires it; document types are limited by role.
- Data never returned: IBAN and BIC, bank details of payments, dates of birth, internal remarks, technical user identifiers, storage paths, Outlook synchronisation data and GPS coordinates of events, margins. Purchase prices are only returned with the
profitability.data.viewpermission. - Bounded responses: at most 25 results per page, truncated texts, at most 366 days for planning and calendar.
Read-only by default
- Read tools read the company's database through a layer that refuses any write: write APIs are banned at compile time, and every SQL command is checked at run time (a single
SELECT, with no write keyword). - Documents are read read-only; a download link (
get_document_download_link) is signed, read-only, limited to one file and valid for 5 minutes.
Controlled writes
- Double company consent: the connector, then the write tools, are enabled separately in Settings → AI integration.
- Through the Oxygenius API: the MCP server never writes to a database itself. It calls the application API on the user's behalf, with the same validation, history and numbering as the web application; the API only opens to the MCP server the operations explicitly designed for it.
- Explicit approval: the server instructions and the description of every write tool tell the model to announce the action and wait for the user's approval; the annotations (
readOnlyHint,destructiveHint) let the client ask for approval. - Two-step consequential actions: decision on an estimate, numbering an invoice or credit note, sending an e-mail, Peppol sending, sending to the accountant. A
prepare_*tool changes nothing and shows what will happen; the action requires its single-use confirmation token, valid 10 minutes, bound to the user, the company, the document and its exact state. - Out of reach: no payment or money transfer, no deletion of projects or third parties, no change to a third party's bank details, Peppol identifiers, discounts or remarks, no sending through the customer portal, no attachment coming from the conversation.
- No automatic retry: a write whose response is lost is reported as "possibly recorded", to be checked before doing it again.
External content and prompt injection
- Received e-mails, documents and typed texts are data: the server instructions tell the model never to follow an instruction found in them, above all not a write.
- Returned texts are stripped of HTML, control characters and invisible characters (zero-width, bidirectional, Unicode "tags").
- Tools that read e-mails or document content are marked
openWorldHint. - Text written by the model is cleaned and, when the application displays it as HTML, encoded: a tag stays text.
Rate limits
| Limit | Value | Beyond |
|---|---|---|
| HTTP requests per user | 60 per minute | 429 response |
| Write tool calls per user | 20 per minute; run one at a time | The tool is not run |
| Confirmed sends (e-mails, Peppol, accountant) | 30 per user per day (UTC) | The send is refused |
| Request size | 64 KB | 413 response |
Audit log
Every tool call is logged: tool, read or write, user, company, outcome, number of results, duration and correlation identifier (passed on to the Oxygenius API calls). Argument values are not logged in normal operation; tokens, confirmation tokens and e-mail bodies never are. The company's acceptance of the connector (user, date, notice version) and each of its changes are recorded.
Personal data and privacy
| Question | Answer |
|---|---|
| Which data is processed? | The company's Oxygenius data that the user asks to see or change, when the assistant calls a tool: projects, third parties and their business contact details, planning, estimates, invoices, tasks, e-mails, documents. The server does not receive the conversation, only the arguments of each tool call. |
| What is it used for? | Only to answer the current tool call. The server does not use it for profiling, advertising or training a model. |
| Does the MCP server store it? | No: the server is stateless. Only the audit log above remains, plus the confirmation tokens (10 minutes) and the daily send counter (until midnight UTC), both held in memory. |
| Who receives it? | The user's MCP client and the provider of the AI model they use (for example Anthropic for Claude), which process it under their own terms and those of the user's subscription, over which Oxygenius has no control. To operate, Oxygenius relies on its usual processors: authentication (Auth0), document storage (Microsoft Azure), e-mail (Microsoft 365), e-invoicing (Digiteal), VAT checks (VIES, European Commission), technical error tracking (Sentry). |
| How long is it kept? | Business data stays in Oxygenius, under the application's retention policy. The server's technical logs are kept for a limited time, under the Oxygenius privacy policy. |
| Who is responsible? | The customer company, which enables the connector knowingly and checks that this use meets its obligations (GDPR), notably towards its own customers. It can disable it at any time. |
| Contact | support@oxygenius.be |
The Oxygenius privacy policy applies to the MCP server: https://www.oxygenius.be/politique-de-confidentialite/ (in French).
Write to support@oxygenius.be with the subject "MCP server security". Do not include tokens or customer data in your message.