How Paperclip handles this
A Paperclip routine fires every Monday on a cron schedule and wakes a reporting agent on a read-only connection. It reads get_score_history for each site, which starts nothing and spends no speed-test allowance, and posts the drops on its task. This week the store's mobile score fell eight points. You open a ticket for the fix and assign it to an agent that can write, one that reaches the Hub through a Paperclip connector sign-in, with an instruction to preview first and wait for your reply.
The agent calls optimize_site with dry_run on and posts the plan on the ticket: the settings xSpeed would turn on at the standard level, and the note that nothing changed. You reply Approved. The real pass applies each setting one at a time, checks the page after each, and undoes anything that breaks it, for up to about two minutes, and the agent posts applied, undone and unfixable, plus the verify URLs. Every step sits on the ticket with its trace, so whoever reviews the change later can see the plan you approved.
The verify step needs a person. The Hub tells the agent to open the verify URLs and check that the pages render with a clean console, or to say which pages to check. The Claude Code agent runs headless, so it lists them on the ticket and someone opens them in a browser. Close the ticket only after that check, and leave the unfixable list on it as the next piece of work.
Set up Paperclip once
Already connected? Skip to the prompts. Alternatives and troubleshooting are on the Paperclip 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 the agent Paperclip runs
For a Claude Code agent, run this on the machine that hosts it, as the same OS user the Paperclip heartbeat runs as. Paperclip's docs say MCP wiring lives at the adapter and runtime layer, and user scope makes the Hub available to every Claude Code agent that user runs.
claude mcp add --transport http --scope user xspeedhub https://app.xspeedcache.com/xspeed/mcp
3Sign in once, or use a token
Start claude interactively as that same user, run /mcp, choose xspeedhub and approve access on the xSpeed Hub page. Claude Code keeps the credentials, so later headless runs need no browser; claude mcp get xspeedhub shows the status. On a host with no browser, add the server with the connection token from Connect AI in the Hub instead, ideally with Read-only everywhere on.
# Interactive sign-in claude /mcp # Or a token, no browser claude mcp add --transport http --scope user xspeedhub https://app.xspeedcache.com/xspeed/mcp \ --header "Authorization: Bearer <your connection token>"
Full Paperclip 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. Paperclip's Claude Code adapter runs headless with dangerouslySkipPermissions on by default, which lets the agent call MCP tools without asking. Set it to false and a tool that needs approval does not run. A server added on the adapter does not pass through Paperclip's action permissions or review queue. A Paperclip connector does: each Hub tool is set to Allowed, Ask first or Off, and because the Hub publishes no annotations Paperclip classifies them by name, which can under-classify, so start every write Off. xSpeed Hub has no confirmation step of its own, so a write that gets past those settings runs as soon as your connection allows writes. For unattended routines, send a connection token with Read-only everywhere on, which makes the Hub refuse every write tool.
What do I ask?
Three prompts written for Paperclip. More for this job are below.
Weekly routine: use xSpeed Hub to read the stored PageSpeed history for every site and post any mobile drop of five points or more. Read only.
Ticket: use xSpeed Hub to preview an optimization for store at the standard level and wait for my reply before applying it.
With xSpeed Hub, post the verify URLs and the unfixable list from the last run on store on this ticket.
What happens, step by step
optimize_site is the Hub tool for "make my site faster". It measures the site, applies the recommended xSpeed settings one at a time, checks the page still renders after each change, undoes anything that breaks it, and reports what it applied, what it undid and what it could not fix. It reaches caching and asset delivery only. Page weight, images hosted on another domain and heavy video are outside it, and the report says so instead of claiming a win.
01
See where the site stands
The agent reads get_score_history for the last stored scores and list_modules for which modules exist on that site and what state they are in. Neither changes anything.
02
Preview with dry_run
optimize_site with dry_run: true returns the plan: each change and its tier, what would be skipped (already on, or needing a higher aggressiveness) and what it cannot fix. Nothing is applied. The tool description tells the agent to prefer this before the first real run on a site you have not optimized before.
03
Apply it
optimize_site without dry_run applies the plan. aggressiveness is safe (removals and server-side changes only), standard (the default) or aggressive, which includes settings known to break some themes and should be opted into deliberately. A run can take up to about 2 minutes per site.
04
Look at the site afterwards
When changes were applied, the result carries verify_urls and a note. The site checks its own HTML and reverts anything that breaks it, but it cannot run JavaScript, so verified: true does not prove the page works in a browser. The Hub tells the agent to open those URLs and check the page renders and the console is clean, or to tell you which URLs to check.
05
Reach for a number only if you named one
target_score (1 to 100, Pro sites only) turns one pass into repeated rounds toward that score, up to max_rounds, with 3 as the most per site. Each round is a real PageSpeed measurement and up to about 2 minutes. The result says why it stopped, and diminishing returns means what is left is outside what caching can reach.
06
Critical CSS on Pro sites
generate_critical_css generates above-the-fold CSS so the first screen can paint without waiting for the full stylesheets. It needs the Pro Critical CSS module on that site, and the site answers with an unknown-tool error without it. Pages are rebuilt with it on their next build, so the agent purges the cache afterwards.
Reference
| Main tool | optimize_site (write): measure, apply one setting at a time, verify, undo what breaks, re-measure |
|---|---|
| Preview | dry_run: true returns the plan and changes nothing |
| aggressiveness | safe, standard (default) or aggressive |
| Every site at once | site: "all" or a list of handles, one result per site, sites run one after another; the Hub's confirm-once instruction does not name optimize_site, so your client's prompt is the gate |
| Time | Up to about 2 minutes per site |
| target_score | 1 to 100; needs Pro on the site; without it the site runs one pass and says so |
| max_rounds | Up to 3 per site; only sent along with target_score; 12 rounds in total per call |
| Wide fan-out with a target | Each site gets fewer rounds, and the result carries a note saying so |
| Result fields | applied, skipped, reverted, unfixable, score, verified, rounds; verify_urls after a run that changed something |
| Already tuned site | Returns "everything is already on" plus what it cannot fix; that is an answer, not a failure |
| Critical CSS | generate_critical_css (write), Pro Critical CSS module required; purge_cache afterwards |
| Read-only connection | optimize_site and generate_critical_css are refused; dry_run is still a write call, so it is refused too. Applies to the connection token with Read-only everywhere on and to a Viewer sign-in, not to other OAuth sign-ins |
Rules worth keeping
- Run dry_run first. optimize_site changes live site settings as soon as your connection allows writes. The tool description tells the agent to preview first on a site that has not been optimized, but that is guidance, so your client's approval prompt and a read-only connection (the connection token with Read-only everywhere on, or a Viewer sign-in) are the real gates.
- Keep aggressive for a deliberate choice. It includes settings known to break some themes, and the agent should say so before using it.
- After a run that changed something, open the pages in verify_urls. The HTML checks cannot run JavaScript, so a page can pass them and still be broken in a browser. Do not call a run safe because verified is true.
- Relay the unfixable list. When a site is already tuned, the honest answer is that caching is done and the rest is page weight, hotlinked images or DOM size. The agent should not retry the tool or claim a gain it did not make.
- Only ask for target_score when you want the site driven to a number. Each round costs a real PageSpeed measurement, and a plain "make it faster" is one pass.
Good to know with Paperclip
The route decides whether waiting for your reply is enforced. Through Paperclip's Connectors page, set optimize_site to Ask first and the call waits in the review queue. The Hub publishes no annotations, so Paperclip classifies its tools by name; check where optimize_site landed, and remember that the preview and the real pass are the same tool, so Off blocks the dry run too. Through the agent's own MCP config, the Claude Code adapter runs headless with dangerouslySkipPermissions on by default and calls tools without asking, and xSpeed Hub has no confirmation step of its own, so the reply on the ticket is a convention. Set dangerouslySkipPermissions to false and a tool that needs approval does not run at all. Keep the routine's agent on a connection token with Read-only everywhere on, which refuses optimize_site, dry run included.
More prompts for this job
They work in any client connected to xSpeed Hub.
Ask xSpeed Hub to preview what optimize_site would change on shop with a safe pass. Do not apply anything.
Use xSpeed Hub to optimize the blog site with the standard pass, then give me the pages to check by hand.
With xSpeed Hub, get docs to a mobile PageSpeed score of 90 if you can, and tell me why it stopped if it does not get there.
Use xSpeed Hub to show me what optimize_site would do on every site, as a dry run, and list which sites have the most left to change.
Use xSpeed Hub to generate critical CSS for shop, then purge the cache so the pages are rebuilt with it.
With xSpeed Hub, optimize shop, then compare the new PageSpeed score with the last stored one.
Using xSpeed Hub, what could optimize_site not fix on the shop site?
Frequently asked questions
Keep going
Raise the PageSpeed score with other agents
More with Paperclip
Documentation
- How to generate Critical CSS
- Performance recommendations
- How to run a speed test
- Free vs Pro: what is included
- How to write prompts for xSpeed Hub
- Paperclip + xSpeed
- Every AI agent that works with xSpeed
From the blog