I Didn't Get MCP, So I Built My Way Through It - Enrique Mogollan

Summary

Enrique Mogollan explains MCP by showing how messages move between an MCP client and server, then demonstrates two applications of the protocol: a Ruby MCP inspector and a Rails-focused server called Coatepec.

MCP Inspector

Mogollan built a Ruby inspector to understand MCP’s underlying behavior. He describes two transport options: stdio for local servers and streamable HTTP for network communication. The earlier server-sent events transport has been deprecated.

MCP messages use JSON-RPC. The inspector demo shows how to configure a server, list its tools, resources, and prompts, inspect their JSON schemas, and invoke them while viewing the raw JSON exchanged.

Context and relevance

Using a night-navigation analogy, Mogollan compares tokens to stars, a model’s context window to the portion of sky currently visible, and attention to the lines that connect relevant stars. MCP adds information to that context, but too much output can make important details harder for an agent to find.

Coatepec for Rails development

Repeatedly running tests required waiting for Rails to start and produced large amounts of output. Coatepec keeps a Rails process warm so an agent can quickly request tasks such as running tests or inspecting routes and models.

Its compact responses reduce irrelevant context. Mogollan says successive versions cut token use by about 90%, in part by trimming JSON and returning concise pass/fail information rather than full error paths.

Performance and scope

Coatepec uses worker processes and database connections, with locking to coordinate work. Mogollan reports that using fork instead of spawn improved performance in his setup.

Available tools include test runs compatible with Minitest, as well as requests for models, routes, and Rails engines. Mogollan emphasizes that Coatepec is for development work—not a Rails console, raw SQL access, or general debugging.

Subagents

Sending work to subagents can help, but their setup and responses also consume tokens in the main context. Mogollan cautions that subagents may not be worthwhile when token budgets are limited.

MCP apps

The talk ends with an example of an MCP app that provides an interactive UI inside a chat window. A server can expose an interface rendered in an iframe, with HTML, styling, and JavaScript that can communicate with MCP tools.

In the demo, a coffee-place card lets the user change a rating or favorite drink and displays an animation. The interface comes from MCP, even though it is embedded in a Rails application.

Demonstrations and guides

1. Inspect MCP server setup, tool, resource, and prompt listings, schemas, and raw client-server JSON.
2. Compare an agent’s verbose Rails test workflow with Coatepec’s faster, more focused response.
3. Embed and interact with an MCP-provided UI in a chat-style interface.

Main speaker/source: Enrique Mogollan, software engineer at Handshake.