About the catalogue
The AI Service Catalogue collects assessments of language model services (LLM) and MCP servers. An organisation in the public sector can use an existing assessment as a starting point, instead of assessing every service on its own.
The catalogue is read-only. It is a proof of concept, and all content is fictional sample data.
Guidance, not approval
An assessment describes what was found at a given date. It does not approve a service for use. The organisation that uses a service remains responsible for that use and must make its own local assessment. That assessment covers, for example, the purpose, the data involved, the users, and whether a data protection impact assessment (DPIA) is required.
What an assessment covers
Each assessment has a summary, findings per area and a list of actions.
Areas for all services
- Privacy
- Legal basis, data processing agreement, and which personal data the service is suited for.
- Data flow and retention
- Where data is processed and stored, how long it is kept, whether it is used to train models, and transfers outside the EEA.
- Information security
- Access control, encryption, logging, vulnerability handling and independent testing.
- AI regulation
- How the service and its intended use relate to the EU AI Act and what the vendor documents.
- Exit options
- Data export, notice period and how hard it is to change to another service.
Additional areas for MCP servers
- Tool access
- Which tools only read data and which can write, delete or send data on behalf of the user.
- Authentication
- How the server identifies users and how narrowly access is scoped.
- Hidden instructions in tool descriptions
- The risk that tool descriptions or returned content contain text that tries to steer the AI assistant.
Status
- Not assessed
- The service is listed, but no assessment has been made. A vendor declaration may be available.
- In review
- An assessment is in progress. Findings and actions can change.
- Assessed
- An assessment is complete. It reflects the service at the date it was last updated.
Severity
Each finding and each action has a severity.
- Low severity
- Minor concern. Normal routines are usually enough.
- Medium severity
- Needs attention or a local measure before use.
- High severity
- Significant concern. Resolve or accept the risk explicitly before use.
Actions and vendor input
Actions are grouped by who should carry them out: the vendor, the organisation, or both. The vendor declaration shows the vendor's own answers to a fixed set of questions. Vendor comments show the vendor's response to the assessment.