From synchronized OData data to bounded MCP tools
OData and MCP solve several problems. OData gives reporting tools a standardized data feed; MCP gives an AI client a small set of described, read-only tools with explicit permissions and limits.
Last content checked: 27 juli 2026
What remains the same
- The source data is pre-synchronized to the tenant database
- Table names, column types, and relationships make up the catalog
- Existing OData access for Power BI remains unchanged
- Both routes read the last synced data and do not call the source API for every query
What MCP adds
- OAuth on behalf of a user instead of a long-lived OData API key
- Permission per connection, source, table and column
- Tool descriptions and typed results for AI clients
- Strict row, byte, time and query limits
- Secure sync status and audit metadata
No random URL or SQL
The model client gets no SQL and no tenant ID. Query's are constructed in a structured manner and every field reference in selection, filter and sorting is checked server-side against the allowlist.
Database side aggregation
PostgreSQL can perform filters and approved aggregations before data goes to the model. As a result, the external assistant receives a small result instead of thousands of individual OData rows. DigiData AI can also use richer approved relationships and metrics than the public MCP contract.
When do you choose what?
- Power BI or Excel: use the existing OData feed
- Remote Assistant: Use the limited MCP connection
- Recurring, controlled business analysis: preferably use DigiData AI
Data flow from source to external assistant
DigiData does not request the source API again for every AI query. The MCP layer only reads pre-synchronized data and applies connection permissions before a result reaches the chosen provider.
- Step 1
Source systems
Synchronize Exact Online, Twinfield, Bouw7 and other linked sources outside the interactive query path.
- Step 2
DigiData tenant database
Synchronized data remains separate per tenant and is read with the current connection rights.
- Step 3
Read-only MCP layer
OAuth, source, table, and column permissions, and bounded tools determine what the connection is allowed to return.
- Step 4
Chosen AI provider
Only the allowed result goes to ChatGPT, Claude or another client and is subject to provider policy.
