A First Look at Optimizely Remote MCP Server for Experimentation

Optimizely just released a Remote MCP Server for Experimentation and I've been trying it out to see what it can do.

If you don't know, MCP (Model Context Protocol) is the "standard" way for AI tools to talk to external services. Optimizely already had a local MCP server, but this new one is hosted directly by them so anyone with an Optimizely login can point Claude (or ChatGPT, or Cursor, or whatever) at it and start chatting with their experimentation data. No local servers or API keys needed.

Setup

Setup was easy enough, although running `claude mcp add --transport http optimizely-exp https://exp.mcp.opal.optimizely.com/mcp` did not work for me, nor did adding the below to my `claude_desktop_config.json` as instructed here. 

{
  "mcpServers": {
    "optimizely": {
      "url": "https://exp.mcp.opal.optimizely.com/mcp"
    }
  }
}

When I launched Claude on Mac I got an "Some MCP servers could not be loaded" error.

At least the fix is simple, we can use mcp-remote for now in `claude_desktop_config.json`

{
  "mcpServers": {
    "optimizely": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://exp.mcp.opal.optimizely.com/mcp"
      ]
    }
  }
}

After that, when I opened Claude Deskop my browser opened to OAuth with my Optimizely credentials and for me to pick a project to connect the MCP with, done.

What can it do?

I started simply with "list my Optimizely projects" and after allowing Claude some actions we got a clean table back. 

Drilled into my Niteco Sandbox to see all experiments. 

Drilled further into a running test called "Metrics Setup" to see how it was doing. 

Claude summarised everything into nicely formatted tables and flagged anomalies I had set up for testing. It pointed out that the traffic split was 0% on Original vs 100% on Variation #1 and that two metrics are performing worse than expected.

Naturally, the next thing was to fix the broken split. I asked Claude to set it back to 50/50, it pulled the entity template, confirmed what it was about to do, and applied the change. 

Then it hit the wall. Changes like these don't go live until you publish them, and even though Claude said they were live already, turns out they weren't. 

I had it double-check.

Seems like for now the MCP is leaning towards the safer side. Let AI read everything and modify the active draft, but not actually do the final click to ship. I can see the logic of this, although hopefully it gets added as a project-level setting to allow AI full control.

Conclusion

For querying, summarising, and inspecting experiments, the remote MCP server seems great. For making changes, it seems like it will be a huge help as well, but you will need to hop into the UI to click Publish one last time. 

I'm looking forward to adding this into our AI workflows to automate manual tasks like reporting and setting up new variant code quickly while giving Claude all the context it needs to get things done correctly first time.