Claude MCP in Practice: How I Connected My SEO Tools
AI summary / tl;dr
- TOPIC: connecting Claude to your own SEO tools via MCP
- SETUP: two in-house tools, one MCP connector each (18 and 16 tools), sign-in via OAuth
- GUARDRAILS: fixed daily limits, source and period on every number, no estimated search volumes, approval before structural changes, no deleting through the connector
- ONE EVENING: 49 → 68 keywords, USD 0.23016 in measurement costs, two bugs in my own code found and fixed
- LIMITS: no ranking results, no forecast figures
With Claude MCP, Claude works directly with your own tools. It reads data, measures keywords and assigns pages without anyone copying spreadsheets back and forth. That is how I connected my two SEO tools to Claude. First I explain what an MCP server is. Then I use one evening in the chat to show what the connection does, which rules keep it safe and which two bugs in my own code surfaced along the way.
What is an MCP server? The Model Context Protocol explained simply
MCP servers in brief: how Claude gets access to your tools
An MCP server is a program that offers tools and data to an AI such as Claude through a fixed protocol. MCP stands for Model Context Protocol, an open standard. Anthropic introduced it and in December 2025 handed it over to the Agentic AI Foundation of the Linux Foundation. The AI calls one of the server's tools, the server runs it and returns the result.
Host, client, server: who talks to whom in Claude MCP
The specification (version of July 28, 2026) distinguishes three roles. The host is the application you work in, for example Claude. It creates a client for every connection. Each client talks to exactly one server, and the server provides the tools. It runs as a process on your computer or as a service on the internet.
One principle of the specification matters for security: a server sees neither the whole conversation nor other servers. The host keeps the history.
Who talks to whom in Claude MCP
- HostClaudekeeps the history
- Clientone per connectiontalks to exactly one server
- Serveroffers toolsdefines what they may do
- Data sourceSearch Console, GA4DataForSEO
Tools, resources, prompts: what an MCP server offers you
A server offers three kinds of building blocks: tools that do something, resources with data to read and prompts as templates. The server announces to the client which of them it supports. In my two servers the tools take center stage: measuring, retrieving, assigning.
Local or on the internet: where your MCP server can run
A server runs either locally on your computer or as a service on the internet. You set up local servers in the desktop app, and they work with your user permissions. Claude reaches a server on the internet through an HTTPS address. If it sits behind a firewall, the firewall has to allow connections from Anthropic. My two servers run on the internet.
Claude connector: how to add your own MCP server to Claude
In Claude, the connection to an MCP server on the internet is called a connector. You add your own server as a custom connector via its address. Sign-in usually runs through OAuth: you log in on the service's own page, and Claude never sees your password. The connection comes from Anthropic's cloud, not from your device.
In its help article on custom connectors, Anthropic points out that such servers are not verified. What a tool may do is defined by the server. You can block individual tools in Claude.
Starting point: two SEO tools, built without formal programming training
I am not a trained programmer. I still build my SEO tools myself, with Claude Code, and run them on my own server. Two of them are connected to Claude.
Keyword Command Center: keywords, clusters and target pages in one place
The Keyword Command Center manages keywords, clusters (groups of related keywords) and target pages, runs analyses and calculates forecasts. It measures search volumes through DataForSEO. Every measured number is stored with its source and data date.
SEO Portal: Search Console and GA4, stored every day
The SEO Portal pulls data from Google Search Console and Google Analytics 4 and stores it daily. Reports work with these stored daily figures. My SEO reporting is built on the portal.
Claude MCP in practice: one connector per SEO tool
Claude MCP servers with OAuth sign-in: 18 and 16 tools
Each tool has its own MCP server, which I added to Claude as a custom connector. Sign-in runs through OAuth. The Keyword Command Center offers Claude 18 tools, the SEO Portal 16. I added eight of the 16 portal tools on November 8, 2026.
What Claude may and may not do through MCP
Claude reads keywords, clusters and Search Console data. It measures new keywords, runs analyses and assigns keywords to target pages and clusters. It only creates new structures once I have approved the plan. Claude cannot delete anything permanently.
Guardrails for Claude MCP: why your rules belong in the tool, not in the prompt
“An instruction in a prompt is a request. A rule in the tool applies to every call.”
Thorsten Litzki
That is why my guardrails live in the server itself and not in an instruction to Claude. They apply regardless of how a request is worded.
Fixed daily limits: how your measurement costs stay predictable
Measurements cost money, and so do AI calls. The Keyword Command Center has fixed daily limits for measurement costs and AI calls. Once a limit is reached, the run stops and says so. The limits sit in the tool's settings and cannot be changed through the connector.
Source and period on every number: how you can check every answer
Every number comes back with its source, tool and data date. Values from Search Console also carry the period of their import. That way I see in every answer where a value comes from and how old it is.
No estimated search volumes: why Claude only gets measured numbers
Language models do not know search volumes. Ask them anyway and you get a number without a measurement. The Keyword Command Center therefore only accepts measured values. Where nothing was measured, there is no volume. A measured value is never replaced by an unmeasured one.
Plan first: structural changes only after approval, deleting only in the interface
Before Claude changes the structure of a project, it shows me a plan. Only after my approval does it create clusters or reassign keywords. Through the connector, Claude can at most archive keywords, and that can be undone. Permanent deletion only happens in the interface.
Two ways into the tools
Claude via the connector
- Request
- Daily limit ✓
- Source and period ✓
- Plan approved ✓
- Tool runs
⊥ separate paths ⊥
Me in the interface
- Sign-in
- Delete permanently
Claude MCP at work: one evening in the chat, from 49 to 68 keywords
In the night from November 7 to 8, 2026, I did the keyword planning for two service pages in the chat: AI Consulting & Automation and Custom Software. Claude worked through both connectors, and I approved the steps.
- matched Search Console data from the portal against the keyword list
- analyzed both service pages: two analyses with 58 ideas
- adopted 19 keywords and measured them with DataForSEO
- checked target pages against the sitemap and the status of the pages
- assigned the new keywords to clusters
| Metric | Value |
|---|---|
| Keywords in the project before | 49 |
| Keywords in the project after | 68 |
| Analyses | 2 |
| Ideas from the analyses | 58 |
| Keywords adopted | 19 |
| DataForSEO measurement costs | USD 0.23016 |
Every step ran through the same guardrails. The measurement stayed within the daily limit, every new number came back with its source and data date, and structural changes only ran after my approval.
- Match1Search Console against keyword list
- Analysis22 service pages · 58 ideas
- Measurement319 keywords · USD 0.23016
- Check4Target pages against sitemap and status
- Clusters549 → 68 keywords in the project
Two bugs in my own code: how MCP access made them visible
Two bugs in my own code surfaced during this work. Both would have distorted numbers without being obvious at first glance.
Bug 1: Search Console values from two periods in one column
The Search Console sync left old values from an earlier file import next to the new ones. As a result, the keyword list mixed values from two periods.
Bug 2: the forecast used outdated positions
The second bug was in the forecast. It calculated with positions that were no longer current.
Fixed and verified: 207 tests and a cross-check through the connector
I confirmed both bugs in the code and fixed them. Afterwards all 207 automated tests passed. Finally, I verified the result through the real connector, the same route Claude uses to retrieve the data.
From finding to cross-check
- 1. Foundwhile working through the connector
- 2. Confirmedin the code
- 3. Fixed207 tests passed
- 4. Verifiedthrough the real connector
The empty day in every total: why your reporting period should end on the last final day
On the same evening I unified the rule for reporting periods in the SEO Portal. Search Console only finalizes the data of the last few days with a delay. Before the change, a still empty day counted toward the reporting period. Over 28 days, the report therefore showed about two percent more impressions on the same data. Now every reporting period ends on the last final day.
Before with an empty day
- 28 days up to the current date
- last day not yet final
Impressions about 2% too many
Now up to the last final day
- 28 days up to the last final day
- same data
Impressions without an empty day
Building an MCP server: three lessons for your own AI integration
Define the rules first, then the tools
The specification has a clear goal: servers should be easy to build. The real work lies in deciding what the AI may do. So I first define which numbers count, what costs money and what should never run through the connector. Only then do I build the tools.
Deliver every number with its source and date
Claude cannot put a number without source and date into context, and I cannot check it. If the tool delivers both, Claude can state them in its answer. That keeps every statement in the chat verifiable.
Use AI access as a test bench for your software
Through the connector, queries run across both tools. The two bugs surfaced during one of these cross-checks. For me, the connector is therefore also a test bench for my own software. I do the same work, defining rules and connecting tools, for companies: in AI consulting and as custom software.
What this Claude MCP case study does not show
This case study measures how I work and what it costs, not rankings. Whether the 19 new keywords bring visitors, there is no data yet. For the path from visit to inquiry, the forecast in the Keyword Command Center currently uses sample values. That is why I do not quote any forecast figures. Customer numbers and revenue are not part of this case study.
Frequently asked questions about Claude MCP and MCP servers
Do I need programming skills for my own MCP server?
Not necessarily. I built my two servers with Claude Code without formal programming training. You do need a clear picture of what the server may do, and secure operation with sign-in, permissions and updates.
Can Claude delete data through MCP?
That is up to the server. If it offers a tool for deleting, Claude can call it. In Claude you can review tool calls before they run or block individual tools. My servers do not offer permanent deletion.
Does MCP only work with Claude?
No. MCP is an open standard. According to Anthropic, ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code support it, among others. Since December 2025 the standard has been with the Agentic AI Foundation under the Linux Foundation.
Can I use the Keyword Command Center or the SEO Portal?
No, neither tool is public. What I offer is the work behind them: a tool or an AI integration built for your company and your data.
Want to connect Claude to your own data? Let's talk about the rules first
If you want to connect Claude to your own data or tools, we start with the rules: what the AI may do, what costs money, which number counts. Book a call
