Oracle and MCP
Let existing AI ask bounded questions and receive compact results with inspectable evidence.
Status: Design. JSON-RPC framing foundations exist. The complete MCP lifecycle, tool surface and real Oracle integration have not been delivered.
Use the model where language matters
An application can use AI to turn a question into a typed query plan. RURAL is designed to check that plan against allowed fields, current permissions, time ranges and work limits before running database operations.
Model inference stays outside the database and outside the event-acceptance path. Ordinary typed database operations must remain usable when AI is unavailable.
A compact tool surface
The proposed Oracle surface has a small set of responsibilities:
Discover the authorized catalog and relevant model or metric definitions.
Prepare or execute a bounded query.
Assemble a compact context packet for a stated purpose.
Expand selected evidence using scoped handles.
Retrieve changes since a known position.
Perform explicitly authorized mutations through a separate transaction operation.
These are tool responsibilities, not final method names or a claim of compatibility with a particular MCP client. The actual client's supported protocol dialect must be pinned and tested before integration is accepted.
Keep evidence close to the answer
An answer should carry its branch/snapshot, source coverage, metric definition and freshness. A short result is useful only when it preserves qualifications that affect the conclusion.
For example, “How many purchases did this product record yesterday?” requires a concrete time zone and interval, the source population, duplicate handling and a statement about missing sources. Unknown data must not become zero just to make the response concise.
Evidence handles let the user or model inspect the records behind an answer without transferring the entire dataset at the start. Current authorization must be checked again when a handle is expanded.
Measure token economy honestly
Filter, aggregate and join in the database. Discover only the relevant schema. Reuse prepared plans, and send a compact unchanged response when polling has no new information.
Evaluation must compare equivalent answers and evidence. Record tool bytes, calls, latency and model-reported input/output/cache usage when available. Byte counts are useful transport measurements; they are not exact billable token counts.
RURAL has not published a token-saving percentage. The target is practical efficiency with preserved correctness and evidence, demonstrated through the real application rather than a synthetic prompt alone.