Oracle Connects NL2SQL to MCP Server, Keeping Database Permissions in the Query Path

Oracle is connecting OCI NL2SQL to its Database Tools MCP Server, offering chatbots and agents a path from a business question to database results. The promise is plain-language access to enterprise data; the constraint is equally central: NL2SQL generates validated SQL, while the MCP server and source database apply configured permissions before a query runs.

That division is the product’s defining design choice. Oracle says customers can connect an approved MCP-compatible client to enterprise data while queries continue to run against the source database, rather than requiring a copy of the underlying database in a separate analytics store. The system is meant to retain configured OCI and database access controls along that path.

Making business language legible

A database schema can identify tables and columns, but it does not necessarily explain the business meaning behind terms such as booked revenue, pipeline coverage or case backlog. Oracle’s semantic store reads approved table and column metadata and can use comments, annotations, synonyms and relationships between database objects to ground a question in the relevant schema.

That context is intended to help the generator find the right objects and joins when an answer spans several related tables. It does not remove the setup work: administrators select approved schemas and create semantic context before end users begin asking questions.

The execution boundary

Oracle separates the credentials used to prepare semantic context from those used for user queries. The enrichment connection can read the configured schema artifacts and metadata, while a distinct, lower-privileged query connection supports end-user execution. That arrangement limits the permissions used on the live query path, even though richer access may have been needed to describe the schema during setup.

The initial configuration

• Create separate Database Tools connections for enrichment and end-user queries.
• Create and enrich a semantic store using the selected connections and approved schemas.
• Configure the MCP server toolset to generate SQL and run authorized source-database queries.
• Connect a client, then validate generated SQL and returned data with questions whose answers are already known.

A narrow starting point

The current release generates Oracle SQL, and Oracle’s published quick start uses Oracle Autonomous AI Database as its source database. Oracle plans support for SQLite and MySQL dialects, but says those future plans and their timing may change. The present release is therefore a governed Oracle-database workflow, not a demonstrated cross-database layer.

Customers can use an Oracle client, a custom MCP-compatible chat or agent client, or an OCI Responses API workflow using MCP Calling. Teams seeking lower-level control can call the Generate SQL MCP tool directly and handle the generated SQL within their own application.

Oracle also supports background processing in the OCI Responses API for questions that produce longer-running database work. It lets a request continue without keeping the original HTTP connection open, with the application retaining a response ID to poll for completion or offer cancellation; it does not make the database query faster. The practical test remains whether generated SQL reflects the organization’s business definitions, joins and permissions before access expands beyond a small set of known questions.