Zylosystems full logo

Digging for Insights in SF #1 — Building MCP Agents with Elasticsearch at Hack Night

Currently working in IT consulting, specializing in No-Code and Low-Code tools like Retool and Bubble to help businesses streamline operations. As a former Amazon software engineer, I have strong experience in building generative AI-based services for image processing. Passionate about developing AI-driven productivity tools to enhance business efficiency and automation.
Nick Park
Hackathon kickoff at Elasticsearch HQ in San Francisco, with the host sharing the event agenda on screen as participants prepare to build MCP Agents.

On the evening of August 19, we joined the MCP Hack Night hosted at Elasticsearch, Inc.’s headquarters in San Francisco. The event focused on building MCP Agents using the open-source project mcp-use, and it gave our team a number of valuable insights.

Why Elasticsearch?

Elasticsearch provides serverless index instances and a Search SDK that let you run search queries with just a few lines of code. During the hackathon, the organizers gave us a sample search.py that queried the public_servers index—for example, searching the description field for “Internet browsing” and returning matching MCP server entries. Those hits could then be mapped to tools and loaded on demand.

This opened up an interesting possibility. Instead of working with a fixed list of tools, we could query for MCP servers that provide fresh and reliable data sources—such as a browser-based document search server or one that returns best practices tailored to a specific audience. By dynamically selecting and connecting these servers, the agent can deliver stronger performance and help documentation stay consistent and context-aware.

The Insight from Code

The core idea turned out to be simple but powerful:

  1. Index MCP server metadata into Elasticsearch as JSON.
  2. Query the index using the Elasticsearch Search SDK.
  3. Parse the search hits and convert them into BaseTool objects.
  4. Register the tools dynamically via the ServerManager.
  5. Allow the MCP Agent to directly invoke the selected tools.

This approach solves a common problem: connecting too many tools to the Server Manager can overload the model. Instead, you can connect only the servers you need at the moment, reducing overhead and improving efficiency.

The Potential with Zylo-docs

Now imagine plugging this workflow directly into Zylo-docs. An AI Agent could call on MCP servers as needed—organizing frequently used terminology, generating release notes on the fly, or pulling in updated references for specific audiences.

At that point, documentation shifts from being a static artifact to becoming a dynamic, interactive tool. This is exactly the kind of future we’re excited to explore, and we plan to bring these ideas into our service.

Curious about where Zylo-docs will go next? Stay tuned to this series—we’ve only just begun digging for insights in SF.