Can you let an AI assistant handle first-line triage for slow queries and locks in a production PostgreSQL cluster without risking a server crash or leaking table rows? In this note, I'll show how the Datasentinel 2026.10 release tackles that problem through its AI Connector module, turning database incident investigation into a safe natural-language conversation.
From Hundreds of Charts to Contextual Dialogue: How AI Connector Works Over MCP and Skills
A stubborn myth surrounds hooking LLMs up to databases: people assume you have to hand over PostgreSQL passwords and let the model fire heavy queries against live tables. When I dug into Datasentinel (the team also maintains the open-source pg_datasentinel extension), what stood out to me was the architectural separation: the AI gets zero network access to the database itself.
The AI Connector module pairs two pieces: a platform MCP server (Model Context Protocol, an open standard for connecting language models to external tools) that serves collected telemetry, and a Skill profile that guides the model's reasoning like an experienced DBA. The connector works with Claude, ChatGPT, Gemini, Microsoft Copilot, and GitHub Copilot from desktop apps, terminals, or IDEs.
Let's trace an end-to-end scenario on an instance named prod-eu-1. In the morning, you open a chat with Claude and ask: "What happened on prod-eu-1 between 08:00 and 10:00, and why did the orders database stall at 9:00?" The assistant pulls metrics over MCP and returns a ranked list of root causes ordered by impact, complete with evidence and a recommended next step. Conversation context sticks around: in a couple of follow-ups, you drill down from a fleet overview to a single query plan, or ask it to generate a dashboard (for example, via Claude Artifacts) or an HTML report.
Practical Diagnostics and the Trade-Offs on prod-eu-1
In day-to-day work on prod-eu-1, that dialogue replaces tedious SQL queries against system catalogs:
- Pre- vs. post-deploy comparison: ask "Compare
prod-eu-1today from 08:00 to 10:00 against yesterday's window—is there a regression?" and the model diffs workload profiles and query plans across both windows; - Live lock and autovacuum analysis: the assistant surfaces active sessions, lock chains (who is blocking whom), missing or redundant indexes, and dead-tuple cleanup lag (
autovacuum lagging).
Checking query plans won't load production either: the new datasentinel_analytics function (requiring Agent v4.0.0+) grants read-only access to planner statistics and builds estimated execution plans without physically running the SQL.
Let's weigh the upside against the cost of this convenience. You save real time on triage and set up daily summaries via revocable tokens. Access is protected by OAuth 2.1 with no database passwords, chat permissions mirror the web UI (live analysis requires a separate privilege), and table rows never leave the database. On the flip side, AI Connector ships as a paid add-on (with a free read-only sandbox), and metrics, SQL text, execution plans, and schema object names do get sent to your chosen LLM.
Release 2026.10 also speeds up compact Workload PDF reports with lock and wait details, adds a documentation MCP server, auto-cleans inactive instances, unifies agentless monitoring filters, and patches dependencies. You can compare this approach with a neighboring tool in our breakdown of Dasha 1.10, which uses MCP for Patroni cluster diagnostics.
To bring AI diagnostics into your workflow safely, keep three rules in mind:
- Delegate initial triage: use chat to spot lock chains, compare post-release windows, and watch
autovacuumlag; - Account for schema sensitivity: remember that query text and table names go to an external model API;
- Verify recommendations before running DDL: lean on
datasentinel_analyticsestimated plans, but validate any new index on write-heavy tables yourself.
That balance lets you hand off the grunt work of PostgreSQL incident analysis to AI while keeping tight control over your production database.

