Your agent can read
your production traces.

PHPRay speaks the Model Context Protocol. Point Claude Code, Claude Desktop, Cursor or any other MCP client at one URL and it can answer "which page got slow after Tuesday's deploy" from real request data — not from guessing at your source code.

Add it in one line

claude mcp add --transport http phpray https://phpray.dev/mcp

Nothing to install and no account needed. Without a token you get the read-only demo account: a WooCommerce store with real recorded traffic, so you can see what the tools return before you decide anything. Rate-limited to 30 requests a minute per address.

Not using Claude Code?

Every MCP client takes the same server. Paste this into your client's MCP configuration — mcp.json in Cursor, settings.json in the VS Code MCP extension, claude_desktop_config.json in Claude Desktop:

{
  "mcpServers": {
    "phpray": {
      "type": "http",
      "url": "https://phpray.dev/mcp"
    }
  }
}

Transport is streamable HTTP over POST; there is no SSE stream, and a GET asking for one is answered with 405 on purpose, which is what the specification asks for. If your client searches a registry instead, we are listed in the official Model Context Protocol registry as dev.phpray/phpray.

Then ask in plain language

• "Which pages on this site are slowest, and what is the p95?"
• "Which SQL fingerprint costs the most time, and who calls it?"
• "Which plugin eats the most time on profiled requests?"
• "Compare the last 24 hours with the previous 24 — did anything get worse?"
• "Show me the slowest request over 800 ms and break it down."

62 slow query fingerprints, showing 5, ordered by total time in the window (avg x calls) — that is what costs the database. Literal values are masked as ?.

 1. total 116 ms · avg 0.2 ms · max 0.7 ms · 501 calls
    SELECT option_name, option_value FROM wp_options WHERE option_name IN (?)
 2. total 107 ms · avg 0.4 ms · max 0.8 ms · 256 calls
    SELECT option_name, option_value FROM wp_options WHERE autoload IN (?)
 3. total 95 ms · avg 0.2 ms · max 0.4 ms · 457 calls
    SELECT DISTINCT t.term_id, tr.object_id FROM wp_terms AS t INNER JOIN wp_term_taxonomy AS tt ON t.term_id = tt.term_id INNER JOIN wp_term_relationships AS tr ON tr.term_taxonomy_id = tt.term_taxonomy_id WHERE tt.taxonomy IN (?) AND tr.object_id IN (?) ORDER B…
 4. total 85 ms · avg 4.7 ms · max 7.6 ms · 18 calls
    UPDATE `wp_options` SET `option_value` = ? WHERE `option_name` = ?
 5. total 83 ms · avg 0.3 ms · max 1.1 ms · 317 calls
    SELECT option_value FROM wp_options WHERE option_name = ? LIMIT ?

Your own sites

Pass a console token and the same tools see your sites instead of the demo, scoped exactly to what that token can read:

claude mcp add --transport http phpray https://phpray.dev/mcp \
  --header "Authorization: Bearer <console token>"

The token comes from the console under Settings. It travels on each request and is never stored by the MCP service — the service holds no database and no credentials of its own. A read-only token cannot start profiling; the server refuses it rather than pretending.

What the tools do

This table is generated from the live endpoint, so it cannot drift away from what the server actually offers. With your own token you also get phpray_profile_url, which turns per-function profiling on for one URL prefix for a few minutes — the only tool that changes anything, and it is hidden on the demo.

Prefer it local?

The same server runs as a binary next to your collector, over stdio, with no cloud at all:

claude mcp add phpray-local -- phpray-mcp

It reads the collector on 127.0.0.1:9191 and needs no account. Binaries for Linux and macOS are in every release; the source is in src/mcp, Apache-2.0, with no third-party dependencies.

What it is not

• It does not read or write your code. It reads recorded request data, nothing else.
• It does not stream: the transport is plain POST with JSON-RPC, no SSE session.
• It cannot change your PHP configuration, restart anything or touch your server.