Claude's Model Context Protocol is here... Let's test it

Model Context Protocol (mCP) is positioning itself as a plug-and-play standard for giving AI assistants reliable access to external data and actions—turning “chat” into something closer to an API-driven workflow. Built by Anthropic (the team behind Claude), mCP defines a common way for an AI client (like Claude Desktop) to discover what a server can provide (resources) and what it can do (tools). The practical payoff is straightforward: instead of custom integrations for every model and every app, developers can expose capabilities through a shared protocol so LLMs can fetch context and trigger server-side operations with fewer brittle glue layers.

The tutorial demonstrates that workflow end-to-end using a storage bucket, a PostgreSQL database, and an existing REST API hosted on Savola (sponsored). The setup mirrors a typical production stack: user-uploaded images live in object storage, profile and relationship data sit in PostgreSQL, and a TypeScript REST service already handles application logic. The mCP server then becomes the bridge that lets Claude pull the right data for prompts and call actions that mutate real application state—such as creating matches or scheduling dates—rather than merely generating text.

At the core of the implementation are two server-defined concepts. “Resources” are read-only data fetches (for example, a database query that returns candidate horses and their relationship status). “Tools” are actions with side effects (for example, writing to the database or invoking an endpoint that creates matches). The code uses Zod schema validation to constrain the shape of inputs and outputs so the model can’t “hallucinate” arbitrary arguments. That matters because mCP requires the model to decide what parameters to pass; typed schemas and descriptions make those decisions more deterministic and reduce malformed calls.

Once the server is defined, the tutorial shows how to run it locally via standard IO for testing, or swap transport mechanisms for deployment (including server-sent events or HTTP). On the client side, Claude Desktop reads a configuration file listing one or more mCP servers and the command needed to start them. After attaching the server, Claude can fetch resources as context—such as querying which horses are single—and then use tools to perform actions like updating the database, with permissions gating write operations.

The broader claim behind the hype is that mCP makes LLM applications more reliable and interoperable, effectively creating “APIs for APIs” so different models and clients can plug into the same capability layer. The transcript also flags the stakes: Anthropic leadership predicts rapid AI-driven coding adoption, but the same automation could introduce serious failure modes if agents mishandle permissions or data. Still, the immediate takeaway is concrete: with mCP, developers can expose their existing data and services through a standardized interface that lets Claude act on real systems with structured, validated inputs.