Table of Contents

Security via Semantic Model MCP Server

This document describes how the Semantic Model MCP Server enforces data access permissions, so users only see the data they are allowed to see.

Important

Data access through MCP services requires that the OAuth connection option is used when connecting to the Semantic Model MCP Server.


Row-level security

The Semantic Model MCP Server itself does not enforce data access permissions. Instead, it relies on the underlying database (for example SQL Server) to handle this. This way, users and access control rules stay in the source system and do not have to be duplicated in the MCP Server.
The MCP Server only validates that the request from the AI agent is authenticated on behalf of a user. It then passes the user id to the database, where RLS policies are enforced when the MCP Server fetches data.

Note

For row-level security to work, make sure the MCP Security Context is configured in the Semantic Model.


How row-level security is enforced

The sections below describe the prerequisite setup and how security is enforced, so users only see data they have access to.

Step 1: MCP connector is added to an AI tool (Claude, Copilot)

As part of the setup required to connect any AI tool like Claude or Copilot to an MCP Server, the tool takes the user through an authentication process when the connection is set up. It typically follows this pattern:

  1. The user adds an MCP connection to the AI tool (for example, a custom connection in Claude).
  2. The AI tool discovers that the MCP Server requires user authentication.
  3. The AI tool takes the user through the authentication process.
  4. The AI tool can now call the MCP Server on behalf of the user.

Step 2: The user makes requests to the MCP server through an AI tool (Claude, Copilot)

The following diagram shows how a user request flows through service layers, and what each layer is responsible for in enforcing security.

%%{init: {"sequence": {"noteAlign": "left"}}}%%
sequenceDiagram
    autonumber
    actor User
    participant Agent as AI Agent<br/>(Claude, Copilot, etc)
    participant MCP as MCP Server
    participant QP as Semantic Model<br/>Query provider
    participant DB as Database with RLS<br/>(row-level security)

    User->>Agent: Natural language question
    Note over Agent: • Translates the user's question into one or more structured JSON requests<br/>• Calls the MCP Server on behalf of the user

    Agent->>MCP: Structured data query + user's access token
    Note over MCP: • Validates the access token<br/>• Extracts the user id from the access token<br/>• Passes the user id to the query engine

    MCP->>QP: Structured data query + user id
    Note over QP: • Translates the structured data query into SQL<br/>• Sets the security context of the db session (user id)

    QP->>DB: SQL query + user id
    Note over DB: • Executes the SQL query against table(s)<br/>• Enforces RLS so only the data the user has access to is returned

    DB-->>QP: Result
    QP-->>MCP: Result
    MCP-->>Agent: Result
    Note over Agent: • Translates the result (rows, columns + metadata) into a user response<br/>• PPTX, Excel file, image, etc<br/>• (HTML) dashboard, chart, table<br/>• Textual response

    Agent-->>User: AI response (file(s) / text)