How Windsurf handles this
You have just finished a round of front-end changes with the agent and want to know if it was worth it. In Windsurf you ask for the PageSpeed picture of the site. The agent calls get_score_history through the xspeedhub server and reads back the stored runs: score, LCP, FCP, CLS, TBT, Speed Index and TTFB, for mobile and desktop. It can tell you which device is weaker and which metric is behind it, without you opening a report.
To measure again, it uses run_speed_test, a PageSpeed Insights run from the Hub that takes up to about two minutes. The Hub saves the result in its own report history, which the dashboard shows on the Reports tab. get_score_history reads the plugin's stored runs, so the new one may not appear there. If you only want a quick check on whether pages are served from the cache, run_benchmark times the home page cached and uncached and does not return a Lighthouse score. Ask the agent to say plainly which tool produced each number.
Windsurf now ships as Devin Desktop. Its main agent is Devin Local, and the earlier Cascade agent remains for existing conversations, so older Windsurf guides may not match what you see. The tools are the same, because they come from the Hub. Ask for results as a short list that names the site, the device and the date of the run, so you can drop it into a ticket.
Set up Windsurf once
Already connected? Skip to the prompts. Alternatives and troubleshooting are on the Windsurf 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 for the Devin Local agent
Run this in a terminal. Devin Local shares its MCP configuration with Devin CLI, and -s user saves the server to ~/.config/devin/mcp_config.json (%APPDATA%\devin\mcp_config.json on Windows) so it applies to every project. Without -s, the command saves to local scope, which is the current project only.
devin mcp add -s user xspeedhub https://app.xspeedcache.com/xspeed/mcp
3Sign in
Run the login command and approve access on the xSpeed Hub page that opens in your browser. There is no token to paste, and the client keeps the credentials it receives. If the server later shows Needs auth in the Devin Local MCP list, click Authenticate, or run devin mcp logout and then login again.
devin mcp login xspeedhub
Full Windsurf 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. The Devin Local agent prompts for approval before it calls any MCP tool. In the prompt you can allow that one tool or every tool on the xspeedhub server, for this session or permanently, and permission rules such as mcp__xspeedhub__get_cache_status allow, deny or always ask for a named tool. Smart mode lets a fast model run what it judges safe without a prompt, and Bypass mode approves everything, so with either of those a write reaches the Hub as soon as your connection allows writes. xSpeed Hub has no confirmation step of its own.
What do I ask?
Three prompts written for Windsurf. More for this job are below.
Using xSpeed Hub, what did the last two PageSpeed runs on shop say for mobile, and what changed between them?
I just shipped the CSS cleanup to blog. Use xSpeed Hub to run a mobile test and compare it with the previous result.
With xSpeed Hub, time shop cached and uncached, then tell me why that is not the same as a PageSpeed score.
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 Windsurf
Devin Local prompts before it calls any MCP tool, and the prompt lets you allow one tool or every tool on the xspeedhub server, for this session or permanently. Allowing the whole server also allows run_speed_test, and each test spends the shared speed-test allowance. Smart mode lets a fast model run what it judges safe without a prompt, and Bypass mode approves everything. Allow get_score_history by name and leave run_speed_test on the prompt. The older Cascade agent can use at most 100 MCP tools at a time across all your servers, and xSpeed adds 29, so switch off servers you are not using if tools seem to be missing.
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 Windsurf
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
- Windsurf + 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