What is an SEO MCP server?

The Model Context Protocol is a standard way to give an AI assistant tools. An SEO MCP server offers tools that return facts about a website: what a crawl found, which pages are broken, what a page’s title is. The assistant calls them instead of guessing.

Do I need to install anything or get an API key?

No. The server is hosted here and the crawl runs here. You add one address to your assistant. The key is the scan’s own private token, the same one in the link to your results.

Can the assistant change my website or start crawls?

No. Every tool is a read of a crawl that has already been made. It cannot start a scan, change the scan, or touch your site.

Which assistants does it work with?

Any client that speaks MCP over Streamable HTTP. Claude Code connects with a single command. Clients that cannot send a custom header can put the token in the address instead.

Nothing. The crawl is free for up to 5,000 pages with no account, and the MCP endpoint is part of every scan. It allows 120 calls a minute per scan.

Model Context Protocol

An SEO MCP server: give your assistant a full site crawl

Ask an AI assistant to audit a website and it will read a page or two and write a confident checklist. What it lacks is the crawl: every page, every link, every status code. Without that it is describing a typical site, not yours.

Enter a domain. When the crawl has run, every scan has its own MCP endpoint. Connect your assistant and it can run all 37 checks, search the stored pages and inspect any address, with nothing to install.

One email when the scan finishes, with the score and every check. No newsletter.

More options: paths, depth, JavaScript rendering, privacy

What to crawlOnly crawl paths matching (one per line)Leave out paths matching (one per line)Maximum link depth from the start pageDrop these query parameters (comma-separated, * allowed)

Drop tracking parameters utm_*, gclid, fbclid and the like, so tagged links are not crawled twice.Stay inside the start folder Start at /blog/ and only pages under /blog/ are crawled.Crawl subdomains of this host Starting at example.com includes shop.example.com and blog.example.com. Starting at blog.example.com includes only that host and hosts beneath it.Lower-case paths Treat /About and /about as the same page.

Include JavaScript rendering checks Each page is also loaded in a real browser and compared with the HTML the server sent. Slower.Keep this scan private No public report page, no place in the rankings or the sitemap.

Crawls up to 5000 pages plus their images, scripts and stylesheets. No account needed: you get a private link to the results.

Connect in three steps

Scan your site with the form above. You are taken to a private results page.
On that page, open Use this scan from an AI assistant (MCP). It shows the endpoint and the command with your token already filled in.
Run the command, or paste the endpoint into your client's MCP settings.

For Claude Code the command has this shape:

claude mcp add --transport http diyseohub-scan https://diyseohub.com/api/mcp \
  --header "Authorization: Bearer YOUR_SCAN_TOKEN"

For a client that takes only an address:

https://diyseohub.com/api/mcp?token=YOUR_SCAN_TOKEN

The transport is Streamable HTTP with plain JSON answers. There is no local process, no Node or Python to install, and no browser running on your machine.

The five tools

Tool	What it returns
scan_overview	The scan in one answer: the site, pages and files fetched, the health score, and a line per check.
list_checks	Every check the scan can run, with its id and what it looks for.
run_check	One check over the whole crawl: its summary, the total, and up to 200 rows.
search_source	Search the stored HTML of every page: contains, does not contain, a regular expression, or a CSS selector that extracts values.
inspect_url	Everything recorded for one address: status, headers, title and headings, links in and out, and how the crawl reached it.

run_check takes any check id: broken links, titles, redirects, hreflang, structured data, near duplicates, orphan pages and the rest.

Things to ask once it is connected

“Give me the five fixes that affect the most pages, with the URLs.”
“Which pages link to the broken URLs? Group them by the template they seem to come from.”
“Find every page whose title is over 60 characters and write a shorter one for each.”
“Search the source for pages missing our tag manager id.”
“Inspect /pricing and tell me why it might not be indexed.”
“Here is the repository for this site. Fix the missing meta descriptions the scan found.”

The last one is where it earns its place. An assistant with your code in one hand and the crawl in the other can go from finding to fix without you copying tables between windows.

Why a hosted crawl suits an assistant

The crawl is finished before the conversation starts. Tools answer from stored data in moments. The assistant is not waiting on a crawl, and your site is not fetched again for every question.

Answers are sized for a context window. A check returns its summary, its total and at most 200 rows, so the assistant knows the size of a problem without reading four thousand lines to learn it.

The facts are the same ones you see. The tools read the same checks as the results page. When the assistant says forty pages have no description, the table on your screen says forty.

Nothing runs on your machine. Local crawler servers need a runtime, a browser and your bandwidth, and they stop when the laptop closes.

Where it stops

One connection is one scan: to compare two sites, add two. It is read-only, and a scan is a snapshot, so after you fix things, scan again and connect the new one. It works with the crawl only: no keyword volumes, rankings or backlink data.

Treat the token as a password. Whoever has it can read the scan, so give it only to assistants you would show the results page to.

For what an assistant can and cannot do for SEO on its own, see Can ChatGPT do SEO?. The crawl behind the endpoint is described on the site crawler page.

Questions people ask

What is an SEO MCP server?: The Model Context Protocol is a standard way to give an AI assistant tools. An SEO MCP server offers tools that return facts about a website: what a crawl found, which pages are broken, what a page's title is. The assistant calls them instead of guessing.
Do I need to install anything or get an API key?: No. The server is hosted here and the crawl runs here. You add one address to your assistant. The key is the scan's own private token, the same one in the link to your results.
Can the assistant change my website or start crawls?: No. Every tool is a read of a crawl that has already been made. It cannot start a scan, change the scan, or touch your site. Fixes are made by you, or by an assistant you have separately given access to your code.
Which assistants does it work with?: Any client that speaks MCP over Streamable HTTP. Claude Code connects with the single command on this page. Clients that cannot send a custom header can put the token in the address instead.
What does it cost?: Nothing. The crawl is free for up to 5,000 pages with no account, and the MCP endpoint is part of every scan. It allows 120 calls a minute per scan, which is far more than a conversation uses.

Crawl a site, then hand it to your assistant

Up to 5,000 pages, 37 checks, one MCP endpoint per scan. Free, no signup.

Scan a site, free