Skip to content
AI Service Catalogue

Sample data. All services, vendors and assessments in this proof of concept are fictional.

Back to the catalogue
MCPAssessed

Mirelle Docs MCP Server

Vendor: Mirelle Software

MCP server that gives an AI assistant access to the Mirelle document archive. The assistant can search, read and create documents.

Last updated

Summary

Mirelle Docs MCP Server can give an AI assistant read access to documents within the user's own permissions. The write tool and the risk of instructions hidden in documents need extra measures before use.

Assessment areas

  • Privacy

    Medium severity

    Documents can contain personal data. The server uses the user's own access rights, so the assistant can only see what the user can see. The connected language model receives document content and needs its own assessment.

  • Data flow and retention

    Low severity

    The server runs in the customer's environment and does not store documents. Tool call logs are kept locally for 14 days.

  • Information security

    Medium severity

    The organisation is responsible for patching and operation of the server. The vendor publishes a software bill of materials for each release.

  • AI regulation

    Low severity

    The MCP server is a connector, not an AI system on its own. The full solution, including the language model, must be considered as a whole.

  • Exit options

    Low severity

    The server can be removed without effect on the archive. The protocol is open, so another MCP server can replace it.

  • Tool access

    High severity

    Read tools: search_documents, read_document. Write tools: create_document. There are no delete or send tools. create_document can create documents in any folder the user can write to, without confirmation.

  • Authentication

    Low severity

    The server uses OAuth 2.1 with the organisation's identity provider. Tokens are short-lived and limited to the signed-in user.

  • Hidden instructions in tool descriptions

    Medium severity

    The tool descriptions were reviewed and contain no hidden instructions. Documents returned by read_document can still contain text that tries to instruct the assistant (indirect prompt injection).

Actions

Actions for the vendor

  • High severity

    Add a setting that asks the user to confirm before create_document runs.

  • Low severity

    Sign the tool descriptions so that changes can be detected.

Actions for the organisation

  • High severity

    Turn off the write tool unless there is a clear need for it.

  • Medium severity

    Assess the connected language model separately.

Shared actions (vendor and organisation)

  • Medium severity

    Test the solution with documents that contain injected instructions before wider use.

Vendor declaration

Answers as given by the vendor. They are not verified unless the assessment says so.

Where is customer data processed and stored?
The MCP server runs in the customer's own environment. Documents stay in the existing Mirelle archive in Norway.
Is any data transferred outside the EEA, and on what legal basis?
The MCP server does not transfer data. The language model that the organisation connects may do so and needs its own assessment.
Is customer data used to train or improve models?
Not applicable. The MCP server does not include a language model.
How long are prompts, outputs and logs kept?
The server does not keep documents. Tool calls are logged locally for 14 days.
Which subprocessors does the service use?
None for the MCP server.
Which security standards or independent audits does the vendor refer to?
Customers can get the source code on request. An external code review was done in 2026.
How can the organisation export its data and end the agreement?
Not applicable. Documents stay in the archive.

Vendor comments

  • Version 2.3 is in testing. We expect to release it in December 2026.

  • A confirmation setting for create_document is planned for version 2.3.