Understand the flow.
Take control.

The PHPRay WordPress plugin gives you the application layer of PHPRay on any host — including shared hosting — without the PHP extension. The current version 0.1.9 is a direct download: phpray-wordpress-0.1.9.zip (Plugins → Add New → Upload Plugin). The address /dl/phpray-wordpress.zip always points at the newest release.

What it captures

• per-request timing at the application layer (request start → shutdown hook; does not see time inside C extensions)
• all SQL via the $wpdb query log, including N+1 grouping
• outbound HTTP via the http_api filters (WP_Http)
• PHP errors/warnings/notices during the request
• components: every active plugin, theme and core area with inclusive vs self time — this is what makes "which plugin slows down /checkout/" a one-glance answer
• level logic identical to the extension (smart thresholds, alert never sampled)

It produces the same JSONL trace record as the extension (trace format) — the collector, dashboard, CLI and PHPRay Cloud consume both without distinction.

Settings

Tools → PHPRay:

• Enabled — the collector can be switched off without deactivating the plugin
• Thresholds — normal/full/alert ms (defaults 200/1000/3000)
• Sample rate — percent of requests that get the deep instrumentation (SQL callers, errors, HTTP calls, per-component breakdown; default 10%)
• Ignore URIs — one prefix per line (e.g. /wp-cron.php, /feed)
• Exclude logged-in admins — skip requests by users with manage_options
• Output mode — see below. Endpoint and Server key apply to the collector and cloud modes.

Output modes

• file — the plugin appends JSONL records to wp-content/uploads/phpray/phpray.log (rotated at 10 MB, 3 files kept). A collector on the same box reads that file ([input] jsonl_path = "<path to phpray.log>" in collector.toml) and gives you the local dashboard, the CLI and the Cloud upload. Works with every plugin and collector version.
• cloud — works from plugin 0.1.1. The plugin talks to PHPRay Cloud directly over the ingest protocol: batches of trace records go to https://app.phpray.dev/v1/traces (envelope with batch_id, sent_at, agent: wp-plugin/0.1.1, every record carrying a site_id derived from the home_url() host and ABSPATH), and per-minute aggregates go to /v1/aggregates, which is what the fleet view, charts and alerts are built from. Paste the server key (prk_…) created in the console. Only normal/full/alert records are shipped as traces; summary requests are counted in the aggregates. Batches are sent at most every 5 s (immediately for alert) after the response has been handed to the web server. The status box in Tools → PHPRay shows the site ID, the last accepted batch, a rejected key (HTTP 401 pauses sending until you change the key and save), the sampling share requested by the console when a site is over quota, and the last error.
• collector — works from collector 0.15 (plugin 0.1.1). The same envelope is POSTed to the local collector at http://127.0.0.1:9191/api/v1/ingest (endpoint details). The records then take the collector's normal path: live dashboard, SQLite, CLI, and the Cloud upload when the collector has [cloud] enabled — with the plugin's site_id, so the site is the same one the plugin would register in cloud mode. Paste a token in Server key only when the collector has an [auth] secret (mint one with phpray token -role user); without a secret the endpoint takes loopback clients only, which is what a plugin on the same box is. Collector 0.14 has no ingest endpoint: with it, keep file mode.

Limits (honest)

• It is a userland library: it cannot see the PHP core, C extensions, or time spent outside the request lifecycle hooks. The extension sees all of that; the plugin sees the application.
• One plugin per site: it traces the site it is installed in, not other sites on the same account or VPS.
• Component time is attributed from SQL and HTTP callers (the plugin or theme file that triggered the query or call); it has no per-function profiler, so self_ms equals incl_ms in the aggregates.
• Records are buffered in a WordPress option (200 records, 60 pending minutes) between batches; on hosts without fastcgi_finish_request() / litespeed_finish_request() the batch is sent after the page output but before the PHP process ends.
• In smart mode it only profiles the sampled share; the summary layer is always on and cheap (a handful of hooks per request; see Overhead).