SEO MCP server: what it is, which ones exist, and what we learned building one
A confession first: the server this article describes picked its own keyword. We asked whether “seo mcp server” was worth writing about, and the verdict came back winnable. About 20 US searches a month is tiny, but the pages ranking for it carry a median of five referring domains — one of the few SEO queries a brand-new website can realistically win. Bigger neighbors failed the same test. Take “ai seo tools”: it pulls 2,400 searches a month, yet the results ranking there have a median of 61 referring domains, and our domain has 15. Live keyword data, a SERP pull, a winnability verdict, all inside one Claude conversation: that is the job an SEO MCP server exists to do.
What is an SEO MCP server?
An SEO MCP server connects an AI assistant like Claude or ChatGPT to live SEO data: search volumes, Google rankings, backlinks, competitor pages, all through the Model Context Protocol. The AI calls its tools mid-conversation, so keyword research, SERP analysis, and content grading run on real data instead of stale training memory.
The “MCP server” half is generic plumbing — an open standard, introduced by Anthropic in November 2024, for handing AI models tools they can discover and call at runtime. Claude Desktop, Claude Code, ChatGPT, and Cursor all speak it now. You add the server to your assistant's config, the tools appear in the chat you already use, and there is no separate dashboard to learn. If the protocol itself is new to you, we wrote a separate plain-English guide to what an MCP server is.
The differences live in the “SEO” half: the search-engine-optimization data itself, and what the server does with it. They decide what your AI assistant can see — which index the data comes from, what a call costs, and whether you get raw API dumps or something the assistant can reason with. Picture the manual version of the same job: a rank tracker in one tab, a keyword dashboard in another, a spreadsheet in a third, you as the copy-paste layer between them. An SEO MCP server collapses those tabs into tool calls the AI makes in real time, with the numbers landing next to the writing they inform.
An SEO MCP server list: what exists in mid-2026
The market splits into three groups: official servers from the big SEO intelligence vendors, open-source community servers on GitHub, and synthesis layers built on licensed data. All are read-only on your website; none will edit anything for you, so they pair naturally with a CMS-side server (more on that below).
| Server | Data behind it | Access and cost shape |
|---|---|---|
| DataForSEO MCP (official) | DataForSEO's own APIs: SERP, keywords, backlinks, on-page | Pay-per-call API account |
| Ahrefs MCP (official) | The Ahrefs index | Requires a paid plan (Lite from $129/mo); shares the plan's monthly unit budget |
| Semrush MCP (official) | Semrush's index | Requires a paid plan (Starter from $199/mo) |
| Open-source servers (GitHub) | Varies: your API keys, free endpoints, or scraping | Free code; data access is the real cost |
| Waild MCP (ours, beta) | DataForSEO APIs plus server-side caching, briefs, and grading | Waitlist; owner instance live in production |
Two caveats. We are in this comparison ourselves, so read our row with the skepticism you would any vendor's. And prices and tool counts move: the figures here were verified in July 2026 and will drift.
Most of Google's own data sits outside these servers. There is an official GA4 MCP server from Google for analytics, while Search Console coverage — your site's actual clicks, impressions, and query-level performance — remains community-built. Practitioners typically stack two to four MCP servers side by side across one AI assistant: an SEO data server, an analytics server for performance data, sometimes a crawler. No single server covers the whole workflow yet. Running several side by side in one assistant is what MCP was designed for.
The GitHub route: open-source SEO MCP servers
The top-ranking result for “seo mcp server” is not a product page. It is a GitHub repository, which tells you who searches this phrase: developers who would rather wire something up than buy something. The open-source route works, with a catch. Official SDKs exist for TypeScript and Python, a minimal working server is a few dozen lines of code, and it is the cheapest way to start. The hard part is never the protocol; it is the data underneath. Rank tracking, search volumes, and backlink indexes cost real money to produce, so free servers either wrap a paid API key you bring yourself, lean on free endpoints that rate-limit and break, or scrape sources that were never meant to be scraped. If you go this route, check how a server gets its numbers before you trust them.
Is there a free SEO MCP server?
The code is free. The data mostly is not. What “free” buys you is a server that shifts the cost somewhere else: into your own API account, into scraping fragility, or into data quality. On the pay-per-call pricing we build on, the numbers are small — a five-keyword overview costs about a cent, and a five-year Google Trends curve costs $0.011. Our entire dogfooding day on July 17, 2026 — keyword research, a campaign calendar over eleven head terms, a run of field tests — came to $0.90 against a $10 daily cap. SEO data priced per call is cheap at small scale. It is not free, and a server that claims otherwise is paying for it somewhere its README should explain.
SEO MCP server vs WebSearch and WordPress MCP servers
Three server types get compared because they show up in the same conversations, and they do different jobs.
A WebSearch MCP server lets the AI assistant run web searches and read pages in real time: good for reading, useless for measuring. It returns content, not metrics — no search volumes, no ranking positions, no backlink counts, no structured SERP features. You cannot learn from a WebSearch tool that a keyword's monthly search volume is 20, or analyze why the third result outranks the fourth.
A WordPress MCP server sits at the opposite end. It is the publish side, writing posts into your CMS through the same chat interface, and it knows nothing about demand or competition.
An SEO MCP server sits between them as the measurement layer: it pulls the volumes, SERPs, and backlinks the other two never touch, and returns verdicts an AI can act on. A workable stack for AI content work chains those pieces — the SEO server to find and validate targets, build briefs, and grade drafts; the WordPress or other CMS server to push the approved draft live. One AI assistant, several servers, each doing the one thing it is for.
What we learned building one
We have run our server in production since mid-2026, on its own marketing. Four lessons from the builder's side apply to any SEO MCP server you connect.
Context is a hard budget. MCP servers inject instructions into the AI's context on connect, and our primary client truncates that injection at 2,048 characters. We measured it: the original 12 KB instruction block silently lost its entire tool catalog past the cut. The fix was structural — a 2,000-character cap enforced in CI, with full per-tool guides served on demand as resources. When you evaluate a server, look at how much it forces into every conversation.
Raw API responses drown a model. A single keyword response from an SEO API runs
to thousands of tokens of nested metadata and null fields. Ours compresses each one to short field
tokens before the assistant sees it; a keyword row arrives as
{"kw":"ai seo tools","vol":2400,"cpc":41.07,"kd":10}, with a legend tool to decode
on demand. That compression is the difference between an assistant that makes forty research calls
in a session and one that stalls after ten.
Caching decides what iteration costs. Every fact fetched from the paid API lands
in an edge database, and repeat lookups come back free, marked with a
_cache flag so the AI knows the provenance of its own evidence. In one recorded
session, the second run of a campaign calendar over the same head terms was answered 100% from
cache, at zero API spend. Agents re-check the same keyword constantly; without a cache, you pay
every time.
Good tools refuse. The most useful thing ours does is say no, with reasons. The marketing-plan store will not create a plan without demand evidence. Paid-channel guidance will not recommend ads for a low-lifetime-value product, and it shows you the business math behind the refusal. The shipping gate blocks publication until coverage, information-gain, and prose checks pass — this article cleared that gate before it went live. A server that only returns data leaves every judgment to the AI; one that encodes judgment catches the mistakes an AI would otherwise make confidently. The full capability tour walks through those gates with real outputs.
How to choose an SEO MCP server
Four questions cover most of the decision. Where does the data come from, a licensed index or scraping? What does a call cost, and does the server warn you before a large fan-out or only on the invoice? Does it cache, so iterating across the same keywords is free? And does it return raw dumps or detailed verdicts? A server that grades a draft against the live SERP, based on the pages that rank, saves you more time than one that hands you ten JSON files and wishes you luck.
The technical questions — transport, authentication, hosting — matter less than they look. Any competent server handles them; data and cost are where the real differences live.
Frequently asked questions
Does Claude need an MCP server for SEO data?
For live data, yes, and the same goes for ChatGPT or any other AI assistant. Training data is months to years stale, with no current search volumes, rankings, or backlink counts. Without a tool connection, a model can only estimate, and it will, plausibly and wrong. An SEO MCP server replaces those guesses with the numbers as they stand today.
What does an SEO MCP server cost to run?
Depends on the pricing model behind it. Vendor-plan servers (Ahrefs, Semrush) cost their subscription: $129–199 a month and up. Pay-per-call servers cost cents per operation — from our own recorded sessions, about a cent for a keyword overview, $0.011 for a Trends curve, $0.36 for a full campaign calendar over eleven terms. A month of steady content work can land under a mid-tier tool subscription; a careless agent fan-out burns through budget just as fast. Cost estimates, usage caps, and caching belong in the server, not in your head.
Can an SEO MCP server generate content for you?
No — not the writing itself, by design. SEO MCP servers provide data (volumes, SERPs, competitor intelligence), and some, ours included, add briefs and grading on top. The AI does the writing; a good server constrains it with real targets, then gates the result. This article is the loop in practice: a brief built from the live SERP, drafted in Claude, graded and gated before publishing. Insights without generation is the honest division of labor.
Are there SEO MCP servers that also publish?
Not among the data servers: Ahrefs, Semrush, and DataForSEO MCPs are read-only, and ours does not touch your CMS either. Publishing is a different server's job — WordPress and other CMS MCP servers exist precisely so an AI assistant can research through one and publish through another.
There is one exception, and it is how this page shipped: a coding agent. If you edit your site with Claude Code or OpenAI Codex and connect ours to it, the same assistant that researched, drafted, and graded the content can commit and deploy it, using the agent's own file and git tools instead of a CMS integration. So the honest answer is not a flat no; it depends on where you work. In a plain chat window, you publish elsewhere. In a coding agent that already reaches your repo, one assistant carries a page from keyword to live URL — this page did exactly that.