xSpeed Cache MCP in 2026: The Tool List Is Not Fixed, and Write Access Is Free
Updated October 2026
Read on 9 October 2026, our MCP tool reference documents 21 tools on the server built into each site, 12 of which write, while xSpeed Hub exposes a separate catalogue of 29. Neither number is a promise. The same page says the connected server’s own tools/list response is authoritative, because the tools available vary with plugin version, registered modules, licence and connection scope.
That single sentence is the most useful thing to know before connecting an agent, and it is the thing a launch announcement never tells you. This article covers what an agent can call, which calls change your site, which server you are actually talking to, and what any of it costs.
| If you want to… | Connect to | Because |
|---|---|---|
| Drive one site from your agent | the plugin’s own MCP server | It ships in the free plugin and needs no account |
| Reach every site from one connection | xSpeed Hub MCP | One endpoint forwards each call to the site you name |
| Know what your agent can do | ask it to list the server’s tools | The catalogue is per-site, not per-product |
| Let an agent change settings | nothing extra | Write scopes are free, and every call is logged |
| Let an agent use your own LLM key | Pro | Bring-your-own-key providers are the paid part |
There Are Two xSpeed MCP Servers, and They Are Not the Same Catalogue
This is the distinction most setup guides skip, including the first version of this one.
The Site MCP is built into the xSpeed Cache plugin on a single WordPress install. You connect to that site and the tools act on that site. The Hub MCP lives at https://app.xspeedcache.com/xspeed/mcp and reaches every site you have connected to xSpeed Hub, which is ours.
| Property | Site MCP | Hub MCP |
|---|---|---|
| Where it runs | the plugin, on one install | app.xspeedcache.com, for your whole fleet |
| Documented tools, 9 Oct 2026 | 21 | 29 |
| Scope of a call | the site you connected | the site named in the call |
| Needs an account | no | yes, a Hub connection |
| Price | free | free |
Our tool reference is explicit about the trap here: most Hub tools share a name with a Site tool, but the site-scoped Hub tools take a site argument and the schemas can differ. An agent that learned purge_cache against one site and then talks to Hub will call it wrong until it re-reads the list.
That is the practical reason to let the agent discover the catalogue rather than hand it one. Which brings us to the part that is genuinely unusual.
No Frozen Tool List, and Why That Is the Right Design
When an agent connects, it asks the server for a tool list and xSpeed Cache answers with the current catalogue. The documentation states the rule plainly: “available tools and arguments vary with plugin version, registered modules, license and connection scope. This is not a frozen full catalog.”
The mechanism is stricter than it sounds. Tools appear only when the command behind them is registered on that site, and when a module is inactive the tool is removed from the catalogue rather than offered as a call that would always fail. An agent asking a disabled object cache to flush itself does not get a confusing error; it never sees flush_object_cache at all.
What this means for a prompt you write once and reuse
Two consequences, and both bite agencies first.
A prompt that names a tool by hand is brittle, because that tool may not exist on the next site. A prompt that asks the agent to list the tools first works everywhere. And a tool catalogue is a capability report: asking an agent what it can do on a site is a faster audit of what is switched on than opening the dashboard.
The honest limitation is that this makes the catalogue unquotable. Any article, including this one, can only tell you what the documentation listed on the day it was read. Our own reference page carried a different count earlier in the year.
Which Calls Read, and Which Ones Change Your Site
Of the 21 documented Site tools, 9 read and 12 write. The split matters more than the total, because the write half is where an agent can cost you an outage.
| Call | What it does | Reversible? |
|---|---|---|
get_cache_status, get_settings, list_modules | Report configuration and cache state | read only |
list_commands | Returns the full command catalogue, which is how discovery works | read only |
run_benchmark, get_score_history | Measure, and read past runs including failures | read only |
purge_cache, flush_object_cache | Clear caches | yes, it refills |
toggle_cache, update_settings | Change configuration | yes, by setting it back |
optimize_site | Applies its own plan one setting at a time | yes, setting by setting |
clean_database | Removes revisions, drafts and transients | no |
clean_database is the one to think about before you grant write scope to an agent you are still learning to trust. It is the only call in the list whose effect a later call cannot undo.
There is also a precise trap worth naming, because it is the obvious thing to try: optimize_site accepts neither a scan_id nor a report URL, and it does not apply a scan report’s findings. It runs its own plan. If you want a scan to drive the fix, read the report and ask for the specific settings.
To see your own site’s split rather than this table, open the MCP Server panel in the dashboard. Our documentation notes it shows the same catalogue the agent sees, including which tools write.
What It Costs, Precisely
Read from Free vs Pro on 9 October 2026, the MCP rows are:
| Capability | Free | Pro |
|---|---|---|
| Token / OAuth authentication | yes | yes |
| Read and write access scopes | yes | yes |
| AI tool catalogue | yes | yes |
| Activity / audit log | yes | yes |
| Hub: connect site, connect AI, control all sites | yes | yes |
| Your own OpenAI / Anthropic / Gemini / OpenRouter key | — | yes |
| Model selection and usage counters | — | yes |
That page’s own FAQ answers the question directly: “No. The MCP server and xSpeed Hub connection ship in the free plugin, so you can drive caching from Claude or ChatGPT without paying anything. Pro adds bring-your-own-key AI providers on top.”
We build xSpeed Cache at WPDeveloper, a Startise company, so read that as a vendor quoting itself and check the page. The part worth stating plainly is the comparison: a free MCP server with write scopes and an audit log is not the norm in this field, which is the ground our MCP head-to-head against WP Rocket’s covers in detail.
Connecting It, in Four Steps
- Open AI → MCP Server in xSpeed Cache settings and click Connect.
- Choose a format: Claude Code (CLI), JSON config, AI prompt, or Remote (OAuth).
- Add the token to your agent’s MCP configuration.
- Ask it to list the server’s tools. That is step four rather than an afterthought, for every reason above.
Our AI agents section documents 24 clients as of October 2026, from Claude Code and ChatGPT to Cursor, Gemini CLI, Zed, Kiro and Windsurf, each with its own connection command and auth method. The per-client walkthroughs start with Claude Code, and the full connection reference is in how to connect an AI agent.
Four things worth asking it first
- A portfolio check-in. Ask for cache status across every connected site and a list of what needs attention. Managing your fleet with xSpeed Hub has the Hub-side setup, and connecting Claude to every site you run walks the five routes to a fleet-wide connection.
- A capability audit. Ask which tools exist on a given site. Missing tools name the inactive modules.
- A measured change. Ask for a benchmark, one setting change, then the benchmark again. Why caching did not fix a slow site explains why the second reading often disappoints.
- A second opinion from outside. Run the free xSpeed Scan on the same URL. It needs no account and reads what a visitor’s browser receives, which is the check an agent connected to your own dashboard cannot make for you.
Frequently Asked Questions
My agent cannot see a tool the documentation lists. Is it broken?
Almost certainly not. Tools appear only when the module behind them is registered on that site, and an inactive module has its tool removed from the catalogue rather than offered as a call that would fail. Ask the agent to list the tools and treat the gaps as a report of what is switched off.
I wrote a prompt naming flush_object_cache and it fails on half my sites. Why?
Because object caching is not active on those sites, so the tool is not in their catalogues. Ask the agent to list tools and act on what it finds rather than naming the call yourself.
I connected to Hub and my old prompts stopped working. What changed?
The schemas. Most Hub tools share a name with a Site tool, but the site-scoped ones take a site argument and the arguments can differ. Re-read the list after switching endpoints.
I gave my agent a scan report URL and asked it to apply the fixes. Nothing happened.
That is documented behaviour rather than a fault. optimize_site accepts neither a scan_id nor a report URL and does not apply a report’s findings; it runs its own plan. Read the report and ask for the specific settings instead.
Do I have to pay for the MCP server?
No. Token and OAuth authentication, read and write scopes, the tool catalogue and the activity log are all in the free plugin, re-read on 9 October 2026. Pro adds bring-your-own-key AI providers, model selection and usage counters.
Can an agent change my site without asking me?
It can call the write tools you granted it, which is why the audit log ships free and why scopes exist. Grant read-only first, read the log after a session, and widen it when you are satisfied.
Which single call should I be most careful about?
clean_database. It removes revisions, auto-drafts and transients, and it is the one write in the catalogue that a later call cannot reverse.
My cache hit rate dropped right after I let an agent tune the site. Did it break something?
Possibly, and the log will say. Check the activity log for what was changed, set the suspect setting back, and re-run the benchmark. Cache hit rate dropped covers the causes that have nothing to do with an agent.
Which AI assistants work with this?
Any client that speaks MCP. Our agents section documents 24 of them as of October 2026 with per-client connection steps, because the auth flow differs even though the protocol does not.
Does this work across several sites without Hub?
Only by connecting to each site in turn. One connection reaching a whole fleet is what Hub is for, and that connection is also free.
How do I see exactly which tools my site exposes?
Ask your agent to list the server’s tools, or open the MCP Server panel in the dashboard. Our documentation notes the panel shows the same catalogue the agent sees, including which tools write.
Is the tool count in this article going to stay right?
No, and that is the article’s point. It is what the documentation listed on 9 October 2026. Ask your own server.
Your First Session, in Five Steps
- Connect one site, then ask the agent to list the server’s tools before anything else.
- Compare that list against the modules you think are active. The gaps are the report.
- Grant read-only scope first, run a portfolio check-in, and read the activity log afterwards.
- Rewrite any prompt that names a tool by hand so it asks for the list instead.
- Run the free xSpeed Scan on the same site, because an agent reading your dashboard cannot see what a visitor’s browser receives.
New to the plugin? Setting up xSpeed Cache on WordPress covers installation and the settings that matter first, and pricing starts at $29 a year, with everything in this article available before you pay anything.
