Add MCP server for AI-assisted plan and Query Store analysis

[erikdarlingdata]

Summary

Add an MCP (Model Context Protocol) server to SQL Performance Studio so users can connect AI assistants (Claude, etc.) and have conversational analysis of query plans and Query Store data — directly from the app.

Motivation

The app already parses plans, runs 30 analysis rules, and connects to Query Store. An MCP server would expose this data through read-only tools, letting AI assistants provide deeper, contextual analysis without users needing to copy/paste plan XML or manually describe their situation.

Design Principles

• Opt-in by default — MCP server is off until the user explicitly enables it
• Strictly read-only — no tool can execute arbitrary SQL or modify anything on the server
• Same data the app already surfaces — plans, analysis results, Query Store metrics, operator trees, warnings, missing indexes

Proposed Tools

Plan Analysis Tools

Tool	Purpose
get_loaded_plans	List currently loaded plan files/tabs with summary info (statement count, warning count, DOP)
get_plan_analysis	Full analysis result for a loaded plan — warnings, missing indexes, operator tree, runtime stats
get_plan_warnings	Warnings only for a loaded plan, with severity and operator context
get_missing_indexes	Missing index suggestions with impact percentages and CREATE statements
get_operator_details	Detailed properties for a specific operator node (by node ID)
get_plan_comparison	Compare two loaded plans — cost, runtime, I/O, memory, wait stats differences
get_parameters	Parameter names, types, compiled vs runtime values for a plan statement
get_wait_stats	Per-statement wait stats from actual plans

Query Store Tools (requires active connection)

Tool	Purpose
get_query_store_top	Top queries by CPU, duration, or reads from Query Store
get_query_store_plan	Plan and analysis for a specific Query Store query
get_query_store_regressed	Queries with regressed performance (plan changes, increased resource usage)

Context Tools

Tool	Purpose
get_server_info	Connected server version, edition, compatibility level
get_database_context	Current database name and context from the plan or connection

What This Does NOT Do

• No arbitrary SQL execution — tools return pre-collected/pre-parsed data only, same as PerformanceMonitor's MCP approach
• No writes — cannot modify server settings, kill sessions, force plans, or change anything
• No file system access — cannot read/write files beyond what the app already does
• No credential exposure — connection details are not surfaced through tools

Implementation Notes

• Follow the same read-only MCP pattern used by PerformanceMonitor
• The MCP server runs locally within the app process
• Tools return the same data structures the app already computes (ParsedPlan, AnalysisResult, QueryStorePlan, etc.)
• Transport: stdio or SSE, depending on what works best with Avalonia's process model

User Experience

1. User enables MCP server in app settings (off by default)
2. App starts the MCP server on a local endpoint
3. User configures their AI assistant to connect to the MCP server
4. AI can now ask about loaded plans, run analysis, query Query Store — all read-only
5. User gets conversational analysis: "Why is this plan slow?", "What index would help here?", "Compare these two plans"