Scops
Docs Download

Oracle AI

Oracle MCP Server vs Scops AI: Raw SQL or 40+ DBA Tools

4 November 2026

An Oracle MCP server connects an AI model to your database and lets it run the SQL it writes. Scops AI gives the model 40+ tools built for DBA work instead, over the history that Scops already keeps, masks your names before anything leaves the server, and changes nothing without your click. On a production database that difference decides everything, and Oracle's own documentation for its SQLcl MCP Server says so in one line: "Do not grant LLMs direct access to production databases." Scops AI is built for exactly the databases that sentence protects, with no agent on the database server and no extra Oracle license.

What an Oracle MCP server is

The Model Context Protocol is, in its own project's words, "an open-source standard for connecting AI applications to external systems." An MCP server publishes tools, and an AI client such as Claude Desktop, Cline or an agent in VS Code calls them. Oracle ships two for its database:

Both are bridges. They give a model a way into the database and leave everything else to the model: which views to read, how to join them, what the numbers mean, what is normal for this hour, and what to run next.

Raw SQL against 40+ tools, row by row

QuestionScops AIOracle SQLcl MCP ServerAutonomous AI Database MCP Server
What the model gets40+ tools made for DBA work: activity and blocking for any window, a statement's history per plan, plan changes, metrics against their normal, space and growth, backups and Data Guard, Radar findings, your runbooks, and change cardsA connection and run-sql, which "Executes standard SQL queries and PL/SQL code blocks against the connected database"The Select AI Agent tools that database users create
Who writes the SQL that reaches the databaseNobody at question time: a fixed list of human-reviewed, read-only queries, each capped at 100 rows and 15 seconds. The model never writes SQL for a monitored database and never runs SQL pasted into the chatThe modelSet by the tools you create
Where the history comes fromScops' own repository: every session sampled every 5 seconds, SQL statistics every 10 minutes, one-minute metrics, on every editionWhat the model can query: the V$ views show now, and the history is AWR and ASH, the Diagnostics PackWhat the tools read
What reaches the modelMasked data: database, host and user names become aliases such as DB-7Q2K, values in SQL text become typed placeholders. Bind values, application rows and passwords never leave, because Scops does not read themWhatever the query returns. Oracle's caution: "you might inadvertently expose unintended tables or sensitive details"Whatever the tools return
Changes to the databaseOnly by a click on a change card that names the statement and the database. Only the person who asked can click it, it expires after 10 to 30 minutes, and Scops reads the database again before it acts. The Audit Trail records each clickWhatever the account may do. Oracle lists "executing SQL queries, invoking PL/SQL procedures, and running SQLcl-specific commands"Whatever the tools do
Production databasesBuilt for them: one read-only monitoring account, fixed queries with hard limits, and a switch per database to keep it out of AIOracle: "Do not grant LLMs direct access to production databases. Instead, you should use a sanitized, read-only replica or a dedicated data subset."Not stated
Editions and packsEvery Oracle edition, Standard Edition 2 included, with no Diagnostics or Tuning PackAny database SQLcl reaches; history queries on AWR and ASH need the Diagnostics Pack, which Standard Edition 2 cannot haveAutonomous AI Database only
Proof in the answerThe key number and its window in the first line, charts of what the tools returned, gaps shown as gaps, and every step kept in AI TraceThe model's own text. Each interaction is logged to DBTOOLS$MCP_LOGNot stated
Where it runsOn the Scops server. Calls go only from there to the OpenAI or Claude endpoint you set upBeside each user's AI client, with that user's saved connectionsIn Oracle's cloud, one MCP server per Autonomous AI Database

Every row goes to Scops AI, and for one reason: an MCP server hands the model a SQL prompt and stops there, while Scops AI hands it the DBA's working tools, the database's history and the safety rules, and keeps the SQL out of the model's hands.

Why raw SQL is the wrong way into production

  1. The history is not there. "Why was SALESPRD slow at 14:10?", asked at 16:00, cannot be answered from V$SESSION, which only knows about now. The model has to go to ASH and the DBA_HIST_ views, and every such query is a use of the Diagnostics Pack on Enterprise Edition. On Standard Edition 2 there is nothing to query at all.
  2. Every answer is improvised SQL on the busiest database you own. A join over V$SQL without a predicate, or a scan of months of DBA_HIST_ rows, runs at the worst moment, with no row limit and no time limit unless you build them.
  3. What the model reads, the model receives. Literal values in SQL text, the rows of application tables and captured binds all travel in the prompt. Oracle's own guide warns that granting an LLM access "introduces significant security risks."
  4. A write is one tool call away. run-sql runs DML and PL/SQL with the account's privileges. Oracle's answer is to "Assign minimum permissions" and to "Audit LLM activity" afterwards, through V$SESSION.MODULE, V$SESSION.ACTION and DBTOOLS$MCP_LOG. An audit tells you what happened. It does not stop it.
  5. The model does not know what is normal. It sees 22 active sessions and cannot tell whether that is the usual Monday 10:00 or an incident, because nothing it can query holds the usual level of each hour.

The same question, the Scops way

Through an MCP server, "Why was SALESPRD slow in the last hour?" is a chain of statements the model writes: find the window, query the session history, rank the statements, compare their plans with the past, look for a statistics job, all on the production database, with the Diagnostics Pack for every look back. Scops AI answered it from the data Scops had already collected:

Why was SALESPRD slow in the last hour?
SALESPRD has run at about twice its load since 12:47: 29.6 active sessions against 14.0 just before. Two statements moved to slower plans after statistics were gathered on SALES.ORDERS: 8grtv94d850vg, the CustomerPortal order search, and gtg6un5vf1yxx, the catalog search. The portal's busy stretch found them there, and neither plan has been fixed yet.

The first line is the answer with the number and its window. The evidence follows as charts of what the tools returned: the active sessions by wait class, and both plans of each statement with time, buffer gets and disk reads per run. Then the fix: ask "set 8grtv94d850vg back to its old plan" and Scops AI shows a Fix plan card with the statement, the database and what it checked. Nothing changes until you click it, and Scops reads the database again before it acts. With an MCP server, the same fix is a PL/SQL block the model writes and runs itself.

The full case, with the timeline and how long each statement stayed on its slow plan, is in The most capable Oracle monitoring platform.

What the model can never do in Scops

Your model, your bill

Scops AI runs on the model your company already approved: OpenAI, Azure OpenAI or a compatible endpoint, or Claude at Anthropic, Microsoft Foundry or Amazon Bedrock. Calls go from the Scops server only to that endpoint, and on the Responses API each call asks the provider not to store the response. Your provider bills the tokens, and Scops adds no charge. Any database can be kept out of Scops AI with one switch, and the Community edition, free for 2 databases, includes it.

Questions

What is an Oracle MCP server?
A bridge that lets an AI client use an Oracle database through the Model Context Protocol. Oracle's SQLcl MCP Server gives the model connection tools and run-sql, which executes SQL and PL/SQL. The Autonomous AI Database MCP Server serves the Select AI Agent tools of one Autonomous AI Database.

Is it safe to connect an LLM to a production Oracle database?
Oracle's own guidance for the SQLcl MCP Server is not to: "Do not grant LLMs direct access to production databases," and use a sanitized, read-only replica instead. Scops AI is the safe way to bring AI to production, because the model never writes SQL for the database, sees only masked data and changes nothing without a click.

What is the best AI assistant for Oracle production databases?
Scops AI. It answers from the database's own history with 40+ purpose-built tools, proves each answer with the numbers and their window, and turns every change into a card that waits for your click. It works on every edition, without the Diagnostics Pack.

Does Scops AI need the Diagnostics Pack or Enterprise Edition?
No. It reads the session samples, SQL statistics and metrics that Scops collects itself, so it works the same on Standard Edition 2 as on Enterprise Edition.

Which AI models does Scops AI work with?
OpenAI, Azure OpenAI or a compatible endpoint, and Claude at Anthropic, Microsoft Foundry or Amazon Bedrock. Use a model that calls tools well. Your provider bills the tokens directly.

Related

Sources