AI agents (MCP)

The profiler is also an MCP server. An AI agent working on your app (Claude Code, Cursor, VS Code) reads the same profiles you see in /_profiler: each request, each query with its line of code, the plan, the index to add, the code that replaces an N+1. So it answers “why is /orders slow?” or “fix the N+1s on this page” from what your app really did, not from a guess.

It is at http://localhost:4983/_profiler/mcp, on the profiler’s own port: the preload opens it for a Node server, and the Vite plugin opens it for React Router. Nothing else to install or run: when your dev server starts, the terminal says where it is.

[query-profiler] http://localhost:4983/_profiler · MCP: http://localhost:4983/_profiler/mcp

If you set QUERY_PROFILER_PORT, or the port was busy and it took the next one, use the port the terminal prints.

Connect your agent

Claude Code:

claude mcp add --transport http query-profiler http://localhost:4983/_profiler/mcp

Cursor, in .cursor/mcp.json:

{ "mcpServers": { "query-profiler": { "url": "http://localhost:4983/_profiler/mcp" } } }

VS Code, in .vscode/mcp.json:

{
  "servers": { "query-profiler": { "type": "http", "url": "http://localhost:4983/_profiler/mcp" } }
}

Any other client that speaks MCP over HTTP takes the same URL.

What the agent can do

Tool	What it gives
list_requests	The recent requests, scripts and tasks, newest first: what ran, its status, queries, time in the database, warnings. Filters by path, kind and warnings
get_request	One whole: its warnings, its Prisma operations with their arguments, each query with its SQL, parameters, rows, time and line of code
fix_n1	For each N+1, the code that replaces the loop, the line to change and the queries it saves
explain_query	The plan, the tables and their indexes, and the findings with their fixes, the @@index line included
try_index	Creates a suggested index in a transaction that is rolled back, and asks the plan again: would the database use it?
analyze_session	The indexes that fix the most across the stored requests, what the rules find, and the tables the time goes to
get_stats	The slowest and most repeated statements, by model and by route

explain_query asks the database for an estimate: nothing runs. With measure: true it runs a read in a transaction that is rolled back, as Measure it does in the page. try_index creates the index the same way, and rolls it back. Every other tool only reads what the profiler already has.

What to ask

• “Look at the last request with warnings and fix its N+1.”
• “Why is /orders slow? Check the plan of its slowest query.”
• “Which indexes would help this app the most?”
• “Run the seed and tell me what it did in the database.” (With the seed run through the preload, see Scripts, workers and jobs.)

How it works

It runs inside the profiler, with the official SDK, @modelcontextprotocol/server. The SDK loads only once an agent connects, never with your app. Each request gets a fresh server, with no session to keep, for clients of the 2026-07-28 protocol and of 2025.

Like the rest of the API, it runs only in development and answers only requests addressed to a local host. A request from a page of another site (an Origin that is not local) is turned away, so a page in your browser cannot call it.

Profiles live in the dev server’s memory: the agent sees the last 100, and nothing once the dev server restarts.