SEO MCP Server: How to Choose One and Check Its Numbers

Connect a Search Console server first, because your own query and click data is the one dataset no vendor can estimate for you. Then add one paid source for one recurring job, such as keyword research or backlinks. Choose by the job, not the brand.

One server per job matters for a practical reason. Every server you connect adds tools the assistant has to choose between, and the next two sections show what that costs in money and in context window.

Cost is set by the data provider's meter, not by MCP. Ahrefs and Semrush draw from the API units on your plan, SE Ranking from credits, and DataForSEO bills what you spend from a $50 minimum payment. Work out monthly requests by dividing your allowance by measured units per request.

API (application programming interface) units are the credits a provider deducts each time software requests data. Ahrefs sets rows per request and monthly units by plan. Semrush's MCP requests use the same unit system as its standard API, with 50,000 units included on SEO Classic Pro and Guru and on Semrush One Starter and Pro+; other plans need a purchased units package. SE Ranking offers a free trial of 100,000 credits and pay-as-you-go from $50 for 250,000 credits, which works out to $0.0002 per credit (our arithmetic: $50 ÷ 250,000).

Our model: monthly requests = monthly units ÷ units per request. Units per request is an assumption you replace with your own measurement: run one representative question and note your usage before and after. The table assumes 500 units per request, an illustrative figure, not an Ahrefs rate.

Ahrefs lists Enterprise as unlimited rows per request with 4 million units or a custom amount. At an assumed 500 units per request, Lite allows 400 requests a month, so measure your own units per question before choosing a plan. We would cap spending per key where the vendor allows it, and prepay small amounts on pay-as-you-go services until you know your real usage.

Each tool a server offers comes with a written description that the client may load into the context window. A server with many tools can use a large share of that window before your question arrives.

The clearest measurement is public. In July 2026 a user measured the DataForSEO server at 89 tools and 42,288 tokens (a token is a small chunk of text, roughly a word fragment). The same issue put the median across 29 other servers at 1,679 tokens, so DataForSEO's list was about 25 times larger (our arithmetic: 42,288 ÷ 1,679 ≈ 25). An earlier user had already reported the error "Your input exceeds the context window of this model" against the same repository.

Two things followed. The reporter corrected the claim that this cost is paid on every turn: Claude Code and Codex defer tool descriptions rather than loading them all, so the cost depends on your client. And DataForSEO responded in August 2026 with version 3.0, a redesign with a much smaller surface. Its README now lists four tools: docs_index, docs_list_sections, docs_search and api_request.

Our rule of thumb: count the tools a server exposes before you connect it, and keep a session to the servers its job needs. Versions change fast enough that a review from six months ago may be wrong.

Most are unvetted. We queried GitHub's repository search for "seo mcp" in repository names and descriptions through its search API on 10 October 2026 and counted 1,066 results. That count includes "skills" (bundles of reusable instructions for an assistant) as well as servers, so treat it as a picture of the ecosystem, not a server census.

The vetted core is small. mcp-gsc had 1,892 stars, an MIT license and a push on 15 September 2026, and DataForSEO's official server had 252 stars, an Apache-2.0 license and a push on 2 October 2026 (GitHub, 10 October 2026).

Even popular servers have shipped quiet faults, and their status changes. As of 10 October 2026, the maintainer of mcp-gsc reported the ignored sort_by option (#54) and the reversed period deltas (#42) fixed in version 0.4.0. The error-flag issue (#53) is still open: the maintainer tightened some error handling in 0.4.0 but is keeping the documented string-error convention, so a failed call can still read as data. An older install may still have the first two faults, which is why checks 2 and 3 below exist. Check 4 applies to every version.

Setup breaks too. Users have reported the tool list not appearing in Claude Desktop (#19), a Windows failure on the Unix command which (#32), and fresh installs failing on 28 July 2026 when the Python MCP SDK (software development kit) released version 2.0.0 (#41). Pin the version you tested.

Our screen: prefer an official vendor server or a project with a recent commit, an open license and maintainers who answer issues. Then run the checks below.

Run five checks on the first day, in this order. Each takes a few minutes, and each has a stated pass. They are written for a Search Console server, and the same method works for any server: compare against the vendor's own interface for the same request.

A local server runs with your user account's privileges, so a malicious or buggy one can read files or send data out. A server holding a key with broad permissions widens the damage if that key leaks. The MCP specification's security guidance covers both.

It also requires clients to show the exact command before a one-click install of a local server, says servers must not accept tokens that were not issued for them, and recommends starting with minimal scopes and widening only when needed.

Our practice, as recommendations beyond the specification:

• Start read-only. A server that can edit pages or publish should be a separate, deliberate decision with a human approving each change.
• Use a revocable key or Open Authorization (OAuth) sign-in per server, never a shared password, and keep keys in environment variables rather than in code.
• Treat anything fetched from the web as text written by strangers. Do not pair a page-fetching tool with write access in the same session.
• Remember that whatever a tool returns enters the conversation and goes to your AI provider.

Search Console tells you who clicked. It does not tell you which AI answers cite your pages or recommend a competitor. Our PageLens.ai MCP server covers that layer: it gives your assistant read-only access to our workspace data through two tools. get_schema lists the read-only tables and columns available for your brand. query_sql runs one read-only SELECT over them. The tables cover visibility, tracked prompts and positions, citations and sources, and competitors.

It is a remote HTTP server at https://pagelens.ai/api/mcp, so there is nothing to install locally. You generate a workspace key in PageLens settings and send it as a Bearer token in the Authorization header. The key is scoped to one workspace, read-only, shown once and revocable at any time.

The server is in private beta. It works with clients that accept a bearer token, such as Claude Code, Cursor, Codex and ChatGPT desktop. Claude and Claude Desktop custom web connectors currently expect OAuth, so it does not connect through the Claude app today.

Because it is read-only, the totals and error checks above apply: compare an answer with our dashboard. For what to ask once connected, see how to track AI search rankings and how to track AI citations over time. If you already pay for Semrush, do you need an answer engine optimization (AEO) tool? covers the overlap.

MCP access is included on Growth ($699 a month), Enterprise ($1,499) and Agency (custom), and is not on Launch ($299), as of October 2026; see pricing. We publish this page, and our server is one of the options in it.

To connect it to your own stack, Book a demo.