Oracle AI

The Best AI Assistant for Oracle DBAs: What It Has to Do, and Why Scops AI Is It

2 November 2026

Judge an AI assistant for Oracle the way a DBA judges it at two in the morning. It must already have the database's history, because V$SESSION only knows about now. It must prove what it says with numbers you can check. And it must not be able to break anything. Scops AI is built to those three tests. The rest of the field, by its own documentation, is built for something else: writing SQL for developers, running whatever statement a model produces, or chatting over a suite's telemetry.

What a DBA needs from an AI, in three tests

It has the history. "Why was SALESPRD slow at 14:10?" asked at 16:00 cannot be answered from live views. The assistant needs session samples, SQL statistics and metrics that someone collected and kept, and on Standard Edition 2 or without the Diagnostics Pack, that someone cannot be Oracle.
It proves the answer. A paragraph that sounds right is worth nothing on a production database. The answer has to carry the key number, its time window and the evidence, and it has to say what it could not measure instead of calling it fine.
It cannot do harm. No generated SQL on the production database, no reading of application rows, no change without a person's click, and nothing confidential in the prompt.

Two more, for the real world: it should work without the Oracle packs, and it should run on the model your company has already approved, with the bill going to your provider.

The field, by its own documentation

Assistant	What it is, in its own words	Has the history	Proves the answer	Cannot do harm
A chatbot in a browser (ChatGPT, Claude, Gemini)	A general model. You paste query output into it.	No. It sees only what you paste.	No. It reasons over your paste and cannot check anything.	Only as safe as what you paste and what you run from its advice.
Oracle Select AI	"Select AI serves as a productivity tool for SQL users and developers"; it "generates, runs, and explains SQL queries from user-provided natural language prompts". Available on Autonomous AI Database and Oracle AI Database 26ai and 19c.	Not built for it. It queries application data, not performance history.	It explains the SQL it wrote.	"The queries generated on your behalf by the user-specified LLM provider will be run in your database."
Oracle SQLcl MCP server	A bridge that lets Claude Desktop, Cline and other clients use SQLcl: tools `connect`, `run-sql`, `run-sqlcl`. "An AI client can use natural language to perform the full range of database operations."	Only what the model queries live.	The model's own output.	Oracle's caution: "Avoid production database access: Do not grant LLMs direct access to production databases." Activity is logged to `DBTOOLS$MCP_LOG`.
Observability copilots (Datadog Bits AI, SolarWinds AI Query Assist, DBmarlin AI Co-pilot)	Bits AI: "AI Agents that chat, investigate, and remediate issues" over Datadog telemetry. DPA's AI Query Assist sends one query and its plan to an LLM for recommendations. DBmarlin's Co-pilot: "Ask questions and get instant recommendations from any screen."	What the suite collected	Varies; none describes checking numbers against tool results	Suite-dependent; none describes a confirm-by-click rule for changes
Scops AI	An agent with 40+ tools over the session history, SQL statistics, metrics, space, health and runbooks that Scops collects, on every Oracle edition	Yes: sessions every 5 seconds, SQL statistics every 10 minutes, kept in the Scops repository	Yes: the number, its window, charts of the tool results, gaps shown as gaps; graded by code against the tool results	Yes: fixed read-only queries, no application rows, masking before the call, changes only by a click on a card

None of the other four passes the three tests. A chatbot has no data, Select AI writes SQL for developers, the SQLcl bridge runs whatever the model produces, and the suite copilots see telemetry, not Oracle session history. None of them is a DBA assistant, and their own documentation does not claim to be.

How Scops AI answers

Ask who held the lock between 09:00 and 10:00 and why they never committed. In the evaluation run described in What We Learned Giving an AI Agent 40+ Oracle Tools, that took one tool call, get_sessions with the blocking aspect and the window, 8 seconds, and the answer named the root blocker, active from 09:05 to 09:49, with 40 sessions waiting on enq: TX - row lock contention. Then the part a DBA cares about: the holder had been active for 320 of 2,695 seconds, about 12%, so the answer said the samples did not prove why it never committed, only that the transaction stayed open while the session sat idle. It gave a likely cause and labelled it as likely.

That shape is the rule. The first line is the answer with the key number and its time window. The evidence follows, charted from what the tools returned. A check that could not run is a gap in the answer, never "no problem". A recommendation comes with the check that confirms the cause and the risk of acting on it.

The tools

The tools follow the way a DBA works through a problem: activity (sessions, blocking episodes, waits, an incident report for a window), SQL (a statement's history per plan, plan changes, the statements touching an object), metrics (one database against its baseline, or two windows against each other), space (tablespaces, segments, growth, datafile history), health (backups, Data Guard, memory, OS log findings, Exadata cells), the estate (rankings, wait breakdowns, overviews), your own runbooks and past incidents, and cards that propose a change. Several tools sit behind one entry point, so the model chooses from a short list and the tool picks the detail. Half the graded questions needed one call; the longest took seven.

Safety by design, not by prompt

Fixed queries only. A human-reviewed list of read-only queries is the only SQL that reaches a monitored database, each capped at 100 rows and 15 seconds. The model never writes SQL for a monitored database and never runs SQL pasted into the chat.
No application data. Bind values, the rows of your tables and passwords are never sent, because Scops does not read them.
Masking before the call leaves. Database, host and user names become aliases; values in SQL text become typed placeholders. The real names come back in the answer. Calls go only to the endpoint you configured, from the Scops server.
Nothing changes without a click. To kill a session, the assistant prepares a card with the statement, the database and what it checked. Only the person who asked can click it, the card expires in 10 to 30 minutes, and Scops reads the session again before acting. Typing "yes, do it" does nothing.
A budget per question: 15 rounds, 1.2 million tokens, 8 minutes. Then it answers with what it has.
No pack views. It does not read AWR, ASH, ADDM or SQL Monitor, so it works the same on Standard Edition 2, and a license switch guards the one card that uses Oracle's packs.

Measured against planted incidents

The question catalog has 200 active questions, 134 rated expert and 65 hard, asked against a test estate with planted incidents so the grader knows what should have been found. Grading mixes a judge model from a different vendor with plain code: a required tool that was never called fails the question, every number in an answer is checked against the tool results, a contradicted fact caps the score at 35, and reporting something unmeasured as clean is among the most severely graded errors. In the latest runs, one model passed 166 of 192 questions (86%) with 98.8% of its numbers found in the tool results; another passed 146 with no high-risk answers at all. We have not found another Oracle assistant whose vendor publishes how it is graded, let alone the numbers.

Your model, your bill

Scops AI runs on OpenAI, Azure OpenAI or a compatible endpoint, or on Claude at Anthropic, Microsoft Foundry or Amazon Bedrock. Your provider bills the tokens; Scops adds no charge, and about 92% of the tokens in a typical run come from the prompt cache. Any database can be kept out of the assistant. The Community edition, free for 2 databases, includes it.

Questions

Can ChatGPT or Claude monitor my Oracle database?
Not by themselves. A chatbot sees only what you paste, has no session history and cannot check its own claims. Connected to a database through a bridge such as Oracle's SQLcl MCP server, it can run SQL, which is exactly what Oracle's documentation warns against on production. Scops AI uses a general model too, but through fixed read-only tools over history that Scops collected.

Is Oracle Select AI a DBA assistant?
No. Oracle describes it as a productivity tool for SQL users and developers: it generates and runs SQL on your data from natural language. It is not built to explain a slow hour or find a root blocker.

Is it safe to connect an LLM to a production Oracle database?
Oracle's own guidance for the SQLcl MCP server says not to grant LLMs direct access to production databases and to use a read-only replica instead. Scops AI avoids the question: the model never writes SQL for a monitored database, and the only changes it can propose wait for a click.

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

Sources

Scops docs: Ask Scops AI, Privacy and data masking; evaluation runs of 28 September and 29 October 2026
Oracle: About Select AI
Oracle: Using Oracle SQLcl MCP Server, SQLcl MCP Server Tools, SQLcl MCP Server in SQL Developer for VS Code
Datadog: Bits AI
SolarWinds: DPA for Oracle (AI Query Assist)
DBmarlin: AI Co-pilot