How Cursor handles this
You are halfway through a theme edit, a font swap or a script cleanup, with the file open next to the chat panel. The question is whether any of it shows up on PageSpeed. In Cursor you ask the agent for the history of the site, and it calls get_score_history to read the stored runs, with score, LCP, FCP, CLS, TBT, Speed Index and TTFB for mobile and desktop. The answer sits beside the code you just changed, which makes it easy to connect a metric to an edit.
After you deploy, run_speed_test gives a fresh PageSpeed Insights reading from the Hub, mobile or desktop, in up to about two minutes. It measures the live site, so an edit that is still only in your editor will not show up. For a faster signal, run_benchmark times the home page cached and uncached and returns no Lighthouse score. Ask Cursor to say which of the two it used each time, so you do not read a response time as a score.
Whether Cursor asks before it uses an MCP tool depends on your Run Mode, set under Settings, Agents, Approvals and Execution. Allowlist runs only what you listed and asks about the rest. Auto-review, the recommended default since Cursor 3.6, runs allowlisted calls and sends the rest to a classifier instead of to you. Run Everything asks nothing. When it does ask, you can expand the call to see its arguments. For this job that choice matters more than the connection itself. Know where results land, too. A run_speed_test result is saved to the Hub's own report history on the dashboard's Reports tab, and get_score_history reads the plugin's stored runs, so the two lists can differ. One call tests one device.
Set up Cursor once
Already connected? Skip to the prompts. Alternatives and troubleshooting are on the Cursor guide.
1Put your sites in xSpeed Hub
Sign in at app.xspeedcache.com with Google or email; the Hub is free and has no site cap. Then connect each WordPress site from its own dashboard: click Connect Hub in the xSpeed Cache top bar, then Connect via xSpeed Hub. Each site needs the free xSpeed Cache plugin.
2Add xSpeed Hub to mcp.json
Add the server to ~/.cursor/mcp.json to have it in every project, or to .cursor/mcp.json inside a project to scope it there. The entry is just a name and a url.
{
"mcpServers": {
"xspeedhub": {
"url": "https://app.xspeedcache.com/xspeed/mcp"
}
}
} 3Enable it and sign in
Open Customize in the sidebar, find xspeedhub and make sure it is switched on. When Cursor starts the OAuth sign-in, approve access on the xSpeed Hub page. There is no token to paste. If the server shows an error, open the Output panel with Cmd+Shift+U and pick MCP Logs.
Customize > xspeedhub (on)
Full Cursor setup, sign-in options and FAQ
Before you send a prompt that changes something
Reads change nothing on your sites, though contact_support emails xSpeed support. Writes run once your connection allows them, so your client's approval prompt and a read-only connection (the token's Read-only everywhere switch, or a Viewer sign-in) are the gates that matter. Cursor decides how to treat an MCP tool call by its Run Mode, and you can expand a call to see its arguments. MCP follows the same Run Modes as terminal commands, set under Settings > Agents > Approvals & Execution. The recommended default, Auto-review, runs allowlisted calls at once and sends the rest to a classifier, not to you. Allowlist runs only what you listed, and with an empty list it asks you every time. Run Everything runs every tool call without asking. xSpeed Hub has no confirmation step of its own, so a tool you allowlist, or any call under Run Everything, writes as soon as your connection allows writes. Auto-review is a model, and Cursor says it is not a security boundary.
What do I ask?
Three prompts written for Cursor. More for this job are below.
Using xSpeed Hub, what is the latest mobile LCP for shop, and how does it compare with the run before it?
I just deployed the font change to blog. Use xSpeed Hub to run a mobile test and tell me which metric moved.
Ask xSpeed Hub for the benchmark history on shop and show it next to my last two settings changes.
What happens, step by step
PageSpeed Insights is Google's Lighthouse audit, and its score and Core Web Vitals are the numbers most people mean by "my speed". Through xSpeed Hub an agent can read the scores a site has already stored, run a new PageSpeed test on mobile or desktop, and read or run a separate benchmark of cached against uncached response time. The three answer different questions, and the Hub tells the agent which to use for each.
01
Read before you test
get_score_history returns the stored PageSpeed Insights and GTmetrix runs for a site: the score per run, with LCP, FCP, CLS, TBT, Speed Index and TTFB, mobile and desktop. It starts nothing. The Hub tells the agent to use it whenever you ask what your score is, or whether it improved, because a new test spends a shared allowance.
02
Run a new test when you need one
run_speed_test runs a real PageSpeed Insights audit on the site from the Hub and returns the score, the Core Web Vitals and the top opportunities. strategy is mobile (the default) or desktop, and one call tests one of them. It can take up to about 2 minutes, so the agent should say it is running before it calls.
03
Test the other device
Mobile and desktop scores differ, often by a lot. For both, the agent makes two calls, one per strategy. A failure is reported with its reason: PageSpeed not configured, quota reached, a timeout, or a site the audit cannot reach.
04
Measure the cache on its own
run_benchmark times the home page with the cache bypassed and again served from cache, and returns the response times. It is a read tool, and it accepts site: "all" for every site. It does not return a Lighthouse score, so the agent must not use it to answer a PageSpeed question.
05
Compare over time
get_benchmark_history returns past benchmark runs, uncached against cached, together with the settings changes on the same timeline, which is how you answer "did that change help?". For PageSpeed scores over time, use get_score_history.
Reference
| Read stored scores | get_score_history (read): PageSpeed Insights and GTmetrix runs, starts nothing |
|---|---|
| Metrics per run | Score, LCP, FCP, CLS, TBT, Speed Index, TTFB, for mobile and desktop |
| Run a new test | run_speed_test (write, uses the shared allowance): one PageSpeed Insights run per call |
| strategy | mobile (default) or desktop |
| Duration | Up to about 2 minutes per run |
| Why a test is a write | It spends a shared PageSpeed allowance and stores a report; it changes nothing on the site |
| Read-only access | A connection token with Read-only everywhere on, or a Viewer member's sign-in: run_speed_test is refused; get_score_history, run_benchmark and get_benchmark_history still work. An OAuth sign-in gets the scopes the client asks for |
| Benchmark | run_benchmark (read): cached against uncached response time for the home page, no Lighthouse score |
| Benchmark history | get_benchmark_history (read): past runs plus settings changes on the same timeline |
| Scheduling | The Hub does not schedule PageSpeed tests itself; its daily pass only reads reports the plugin already stored |
| Where a new test is stored | In the Hub's own report history, the one the Hub dashboard shows |
Rules worth keeping
- Check the history first. Ask for stored scores before starting a test, and run a new one only when you want a fresh measurement or something changed.
- Never answer a PageSpeed or Lighthouse question with run_benchmark. It times your own cache and carries no score.
- run_speed_test changes nothing on your site, but the Hub classes it a write because it spends a shared allowance. A connection token with Read-only everywhere on, or a Viewer sign-in, refuses it. An OAuth sign-in gets the scopes the client asks for, so there your client's approval prompt is the gate.
- Lab scores move a few points between runs of the same page. The agent should compare like with like, same site, same strategy, and not report one run's difference as a trend.
- A new test does not run on a timer. If you want a regular test, the agent has to be started by something else, such as a scheduled job in your client, and a scheduled run that spends the allowance should be one you are comfortable leaving unattended.
Good to know with Cursor
Under Run Everything, Cursor runs every MCP tool without asking, run_speed_test included, and Auto-review hands calls that are not allowlisted to a classifier that Cursor says is not a security boundary. A chat that says "check the score after each change" can then start a test after every edit, and each test spends the shared speed-test allowance. Switch Run Mode to Allowlist (Settings > Agents > Approvals & Execution), list the read tools, and leave run_speed_test off so it asks; under the default Auto-review, unlisted calls go to a classifier. Cursor says Cloud Agents never ask for approval, so give a Cloud Agent the connection token with Read-only everywhere on.
More prompts for this job
They work in any client connected to xSpeed Hub.
Using xSpeed Hub, what is the mobile PageSpeed score for shop? Check the stored history before you run anything.
Use xSpeed Hub to run a new mobile PageSpeed test on blog and tell me the score, LCP and the top opportunities.
With xSpeed Hub, test shop on desktop and mobile and tell me which one is worse.
Using xSpeed Hub, has the PageSpeed score on docs improved over the last few runs?
Use xSpeed Hub to benchmark the cache on every site and tell me which one saves the least time.
Did turning on minification on shop help? Use xSpeed Hub to compare the benchmark history around that change.
My PageSpeed test failed on staging. Use xSpeed Hub to tell me why.
Frequently asked questions
Keep going
Track PageSpeed scores with other agents
More with Cursor
Documentation
- How to run a speed test
- PageSpeed Insights integration
- Running speed tests in xSpeed Hub
- How to run a PageSpeed audit
- How to write prompts for xSpeed Hub
- Cursor + xSpeed
- Every AI agent that works with xSpeed
From the blog
- Reading a PageSpeed Report in 2026: 30% of the Score Is a Metric Google Does Not Rank You On
- 45% of the WordPress Sites Our Scan Calls Clean Are Failing Google in 2026: Lab Scores Against Field Data
- The Same Page Scored 61 and 73 Twelve Minutes Apart: How to Benchmark WordPress Caching Plugins in 2026