Claude SEO MCP: how to connect Claude to your SEO data, step by step

What is a Claude SEO MCP server?

A Claude SEO MCP server is an SEO tool that speaks the Model Context Protocol, so Claude can call its functions and read its data while it answers you. The protocol's own documentation describes MCP as an open standard for connecting AI applications to external systems, and compares it to a USB-C port for AI applications: one plug shape, many devices.

In MCP terms, Claude Code, Claude Desktop or Cursor is the host. For each server you add, the host runs a client that keeps a connection to that server. A server can offer three kinds of things:

• Tools: functions the model can call, such as “get rankings for this site”.
• Resources: data the client can read as context, such as a file or a record.
• Prompts: reusable templates, such as a ready-made weekly review.

Remote SEO servers usually use the Streamable HTTP transport, which MCP's architecture page says supports standard HTTP authentication such as bearer tokens, API keys and custom headers. That is why most SEO MCP setups come down to two values: a URL and a token.

Without a server, Claude can explain SEO concepts but cannot see your site's clicks, positions or crawl errors. With one, “which pages lost clicks this week?” becomes a question it can actually answer, with the numbers taken from your account.

What you need before you start

You need an MCP-capable client, an account with an SEO tool that runs an MCP server, and a site connected to that tool. The server can only return what the tool has.

• A client. Claude Code (the terminal and IDE tool), the Claude desktop app or claude.ai, or Cursor. Other MCP clients work the same way if they support remote HTTP servers.
• An SEO MCP server and a token. For MrKeyword, the endpoint is https://app.mrkeyword.io/mcp, and tokens are created in Settings → Connections → API & MCP Tokens. The Free plan is enough to connect, and connecting an assistant has no separate charge.
• Data in the account. Connect Google Search Console (and Bing Webmaster Tools if you use it) to the workspace first. A brand-new workspace with no synced data will answer most questions with “no data yet”, which is the honest answer.

How to connect Claude to an SEO MCP server, step by step

Create a token, add the server to your client with that token in an Authorization header, then confirm the tools loaded. The steps below use MrKeyword's server; swap in your own tool's URL and token format if you use a different one.

1. Create a read-only token

   In MrKeyword, open Settings → Connections → API & MCP Tokens, give the token a label (for example “Claude Code, laptop”) and create it. Leave “Allow this token to make changes” unticked. The value starts with mk_ and is shown once; MrKeyword stores only a hash of it, so copy it somewhere safe now.

2. Add the server to your client

   Use the command or configuration for your client in the next section. Each one needs the same two things: the URL https://app.mrkeyword.io/mcp and the header Authorization: Bearer mk_….

3. Check that it connected

   In Claude Code, run /mcp to see the server's status. In Claude Desktop and Cursor, open the client's tools or connector list. You should see tools such as list_sites, get_rank_summary and get_site_audit.

4. Ask a first question

   Start with “List my sites in MrKeyword”, which calls list_sites. Every other site tool needs a site id from that list, so Claude usually calls it first anyway.

Claude Code: one command

In Claude Code, add a remote SEO MCP server with claude mcp add --transport http, the server name, the URL and a --header carrying the token. Claude Code's documentation recommends the HTTP transport for remote servers and lists SSE as deprecated.

claude mcp add --transport http mrkeyword \
  https://app.mrkeyword.io/mcp \
  --header "Authorization: Bearer mk_your_token"

Two options are worth knowing. Scope: by default a server is added at local scope, for the current project only, stored in ~/.claude.json. --scope user makes it available in every project. --scope project writes it to .mcp.json in the repository so teammates get it too, which is a bad place for a personal token. Management: claude mcp list shows configured servers and their health, and claude mcp remove mrkeyword takes one out.

Claude Desktop and claude.ai: a custom connector

In the Claude apps, a remote MCP server is added as a custom connector: Customize → Connectors, then + Add → Add custom connector, with a name and the server URL. Anthropic's help centre says custom connectors are available on Free, Pro, Max, Team and Enterprise plans, with Free users limited to one, and that the feature is in beta.

• Authentication. Besides OAuth, Anthropic's article says connectors can send fixed credentials, such as a bearer token, in request headers, stored encrypted. For MrKeyword that header is Authorization: Bearer mk_….
• Where the connection comes from. Custom connectors connect from Anthropic's cloud, not from your computer, so the server must be reachable over the public internet. MrKeyword's endpoint is.
• Team and Enterprise. An owner adds the connector in Organization settings → Connectors, and members connect it. Anthropic notes that with fixed credentials every user shares that credential's access, so on a team plan use a read-only token for one workspace.

Claude Desktop also reads a local claude_desktop_config.json (under Settings → Developer → Edit Config). Anthropic's documentation describes that file only for local servers that Claude Desktop starts on your machine, so MrKeyword, a remote server, goes in through the connector route above. The developer reference lists the exact values to enter.

Cursor and other MCP clients

In Cursor, add the server to .cursor/mcp.json in a project, or ~/.cursor/mcp.json for every project, with a url and a headers object. Cursor's documentation supports environment variable interpolation in url and headers, which keeps the token out of the file.

{
  "mcpServers": {
    "mrkeyword": {
      "url": "https://app.mrkeyword.io/mcp",
      "headers": { "Authorization": "Bearer ${env:MRKEYWORD_TOKEN}" }
    }
  }
}

Any other client, or a plain script, can talk to the same endpoint: it is JSON-RPC 2.0 over a single HTTP POST. A tools/list request with your token returns the exact tools your token can use, which is the authority if a page, including this one, disagrees.

What can you ask Claude once the SEO MCP server is connected?

Ask questions about your own site's search performance, rankings and technical health; Claude picks the tools, calls them and writes the answer from what they return. These are example prompts with the MrKeyword tools they usually lead to. They are examples, not recorded answers.

The server also offers four MCP prompts, ready-made workflows a client can show as commands: weekly_seo_review (what moved and five actions for the week), fix_my_top_issues (the latest crawl, grouped by category with fix steps), find_keyword_opportunities (near-miss Search Console queries turned into a prioritised list) and client_report (a client-ready report from rankings, queries and the latest audit). Claude Code lists server prompts as slash commands, so with the server named mrkeyword, typing /mcp__mrkeyword__weekly_seo_review runs the weekly review.

Three habits make the answers better. Name the site and the period (“last 28 days”). Ask Claude to cite the tool and date range behind every number. And ask for a short list of actions, not a dump of rows.

Security and permissions: what to check before you connect

Treat an MCP token like a password, start read-only, connect only servers you trust, and review each tool call before you approve it. Anthropic's help centre warns that custom connectors are not verified by Anthropic, that malicious servers may use prompt injection, and that server behaviour can change without notice. Claude Code's documentation says servers that fetch external content can expose you to prompt injection risk.

• Least privilege. MCP's security guidance recommends minimal, low-risk scopes first. In MrKeyword, every token gets the 20 read tools; only a token created with “Allow this token to make changes” (the actions scope) can call the 5 action tools, and an existing token cannot gain it later.
• Approvals. Claude Code shows a permission prompt for MCP tool calls unless a rule allows them. Its permission rules name MCP tools as mcp__<server>__<tool>, so mcp__mrkeyword__get_* can auto-approve the read tools while action tools still ask. Cursor also asks for approval before running an MCP tool by default.
• Keep tokens out of repositories. Do not commit a token in .mcp.json or a project mcp.json. Use user scope or environment variables.
• Revoke and rotate. One token per client makes it easy to revoke the right one. MrKeyword tokens stop working on the next call after revocation, when an optional expiry passes, or when the person who created them is removed from the workspace.

The server allows 120 calls a minute per token; going over returns HTTP 429 with a Retry-After header. Starting a crawl or tracking keywords uses the same plan allowances as doing it in the app.

Troubleshooting a Claude SEO MCP connection

Most failures are a wrong header, a missing transport flag or a site id that belongs to another workspace.

What MrKeyword's MCP server does not do

It reads one workspace's first-party data and queues a few in-app jobs; it is not an autonomous SEO agent and it has no web-scale index.

• No live SERP ranks. Positions are period averages from Search Console or Bing Webmaster Tools, never a scrape of the results page. The server's own instructions tell the model to say so.
• No backlink or competitor data. There are no tools for other sites' links, rankings or traffic, and the built-in prompts tell Claude not to discuss them.
• No market-wide keyword volume by default. Keyword volume tools appear only when live Keyword Planner lookups are switched on, and each call spends lookup credits.
• No changes to your site. No tool edits pages, publishes content or touches Google settings. The judgement and the publishing stay with you.
• No resources, sampling or SSE stream. The server advertises tools and prompts only, and answers one JSON-RPC POST per call.

Questions and answers