How I Built an MCP Server for agents to use my Algo Update list

For twenty years, the value of a website was the information on its pages. But we are now entering an era where generic content is no longer king. The real value now lies in proprietary data and insight, interactive tools, and making your knowledge actionable for both humans and AI agents.

With this in mind, I started thinking about how I could transform my list of Google algorithm & AI related changes into a tool that you could use with your AI agent. The best way to do this (for now at least) is using MCP - Model Context Protocol. I am really pleased with how well it works!

MCP is a way to enable AI agents to use your tools. You run an MCP server that tells AI agents what tools you have available and how to use them.

In creating this, I found the process was clunky. The end user needs to install the MCP in their LLM tool (Claude, ChatGPT, Cursor, etc.) and this is quite technical. Eventually I will make this page WebMCP enabled which will make things a little easier I think.

I am not an expert developer. And I think this is the exciting thing about this. If I can set up an MCP server with the help of AI, then you can too!

Where to find my MCP server

You can find the Model Context Protocol (MCP) server and instructions on how to use it here: algo.mariehaynes.com.

The algorithm update page is here: Marie's Google Algo Update & AI Changes List

You can see anonymized usage stats here: https://algo.mariehaynes.com/stats

This system uses OKF - the Open Knowledge Format to make agent lookup much more accurate. Now, the system uses OKF, but it's not a true OKF tool where the entirety of the OKF bundle is given to the agent. Rather, in this first release of the tool, the relevant knowledge from the OKF is distilled into targeted tool calls. In the future I am going to learn more about truly making this an OKF tool where the agent can traverse through every data point, but not just yet.

What you can do with it

Instead of copying and pasting hundreds of update dates into a prompt, an MCP server gives your AI agent direct "tools" it can call on demand. When you ask your LLM:

"Why did my organic search traffic drop between August 15 and September 2, 2026?"

Your AI agent automatically queries the MCP server in the background, pulls verified dates and observations from my archive, and uses that ground truth to analyze your site.

When this happens, there is zero LLM token cost on my side. The server purely provides structured data; your client agent does the thinking. The tokens are used within your LLM's plan - like your Claude, ChatGPT or Antigravity subscriptions.

Also, this is much more accurate than simply stuffing the full algorithm update list into an LLMs context window as the tools I've defined make it easier for the LLM to find the data it needs.

Once you've connected the MCP you have all my algo update info in Claude:

If your LLM is good at image analysis you can upload a screenshot of your GSC clicks. ChatGPT's new GPT-6 Astra is quite impressive here:

Here's the geeky info on how I made this, with extensive help from Google's Antigravity.

First, we improved the algo update page design and made it an OKF based system

You can see the old version of this page here. After years of updates (since 2011!), it was quite unorganized! And also I thought it looked boring.

I fired up Antigravity and tapped the mic and rambled about what I wanted. Building with Antigravity is mostly just me conversing like I would with a human.

It went something like this,

Hey Gemini. This is my algo update page and there's a few things I want to do with it. First, I want to be able to update this article automatically as I add relevant stories to my news OKF. Look in our project called OKF-News-System. I want to add something that allows me to just click a button or check a box to say add this to the algo update list....

I went on for many minutes describing what I wanted.

We added an "add to algo update list" button to my News OKF system. This system ingests articles from the SEO & AI News feed in my community and puts them in a dashboard that lets me decide which stories go to clients and into my newsletter.

It's a damn cool system.

When I'm ready to sync, I click a button and the system connects with the WordPress API and publishes to my site.

How cool is this? Now, instead of a boring design we have beautiful cards, tags and the ability to search. So much better!

One of the coolest things in this process was when Antigravity went through 15 years of updates and converted them into OKF format. I had Antigravity make me a dashboard where I could see each entry and suggest changes. Now, every one of my entries exists as a Markdown file with YAML frontmatter. (If this sounds confusing, there's more here on OKF and How I built my OKF brain.)

This process was not perfect because some of my older entries were not well structured. I had to manually read through every entry and suggest corrections for quite a few.

Now we are all set to expose this to agents.

Key Technical Decisions & Architecture

When designing this, I wanted it to be fast, reliable, privacy-conscious, and universally compatible. Full disclosure: I didn't know how to code any of this myself, but  Antigravity walked me through it step-by-step.

In fact, I asked Antigravity to write this part of the article for those who want to better understand the technical parts of what happened here.

Here are the key decisions behind the build:

1. Broad Compatibility (Local & Cloud Transports)
   The MCP spec is evolving fast. We set up the server to support both modern streaming HTTP and Server-Sent Events (SSE) so it seamlessly connects whether someone uses a desktop client (like Claude Desktop via a gateway) or cloud-based tools.

2. The Guardrail: Separating Verified Facts from AI Advice
   My biggest concern was preventing hallucinations. I didn’t want an LLM inventing bad advice and claiming I said it! Every tool instruction strictly forces the AI to output two separate sections:

   • 📊 Verified Update Intelligence: Confirmed dates, official rollouts, and my verified analysis.
   • 💡 AI Strategic Recommendations: The LLM’s own suggestions, clearly labeled as AI advice.

3. Four Core Tools for AI Agents
   The server gives the AI four simple capabilities to query my database:

   • get_latest_updates: Fetches recent major algorithm and AI search shifts.
   • get_updates_by_date_range: The workhorse for traffic drops—finds everything that rolled out within a specific timeframe.
   • search_updates: Full-text search across 590+ updates going back to 2012.
   • get_all_categories: Lets the AI explore updates by category or platform.

4. Strict Privacy by Default
   Many SEOs analyze confidential client data. The server never logs user queries, search terms, URLs, or client identifiers. It only tracks aggregate tool usage counts so I can monitor server health (which you can view live on my open dashboard at algo.mariehaynes.com/stats).

5. Serverless & Lightweight
   It’s hosted on Google Cloud Run, meaning it scales down to zero when idle (costing virtually nothing) and spins up instantly when an AI client requests data.

I had Antigravity publish everything to a GitHub Repository. I’m not a professional developer, but if you’re curious about how the code looks or want inspiration for your own MCP server, feel free to explore it!

How You Can Connect It in 2 Minutes

You can connect this directly to Claude Desktop right now:

1. Open Claude Desktop → Settings (Cmd + , on Mac or Ctrl + , on Windows) → Developer tab → click Edit Config.
2. Add the marie-haynes-algo server to your claude_desktop_config.json:

{
  "mcpServers": {
    "marie-haynes-algo": {
      "command": "npx",
      "args": [
        "-y",
        "supergateway",
        "--sse",
        "https://algo.mariehaynes.com/sse",
        "--logLevel",
        "none"
      ]
    }
  }
}

3. Restart Claude Desktop.

If this is confusing, there are more detailed instructions and even a prompt you can use to get your LLM to help you as well as instructions for other systems here.

Once connected, try asking:

"Did a Google algorithm update roll out between [Start Date] and [End Date]? Please check Marie Haynes' update archive."

There are other sample queries you can try on the MCP page.

Check out how good Antigravity is at using the tool

Where We Go From Here

I've been trying to build an automated traffic drop assessment tool for so many years now - even before we had the help of AI. This is the first time I've been happy with the results.

I want to eventually connect this tool to my OKF brain so that it doesn't just have dates and quick observations of what changed in Search but also has loads of the knowledge that has been in my consulting reports over the years. I really think it could do as good a job as me at producing a preliminary assessment of whether a site has been impacted by Google updates or other search related changes.

Next I will build in WebMCP capability so your agents can use this information right on my website and not require you to install an MCP server.

I recently heard someone say this and it really impacted me,

"I don't want to use your agent. I want to use my agent to use your tools." 

We will increasingly use our own AI systems to get things done. The real value is in curated knowledge and data. It is no longer information that makes a website valuable, but rather, providing value that helps the user accomplish a goal.

I hope you find this tool useful!

➡Instructions for using the MCP

If you liked this, you'll love my newsletter:

Or, join us in the Search Bar for up to the minute news on Search and AI.