← Voltar aos guias

What Is MCP? Build a Useful AI Tool Example

What Is the Model Context Protocol?

The Model Context Protocol, usually called MCP, is an open protocol for connecting AI applications to external tools and data sources. Instead of manually building a different integration for every model, agent, database, or internal service, developers can expose capabilities through a consistent MCP interface.

Think of MCP as a standard connector between an AI host and useful systems. A host such as Claude Code, Codex, or a custom agent can discover available tools, read their descriptions, send structured arguments, and receive structured results. The model decides when a tool is relevant; the MCP server performs the actual work with controlled permissions.

MCP is not a model provider and does not replace an LLM API. It solves a separate problem: giving models safe, reusable access to context and actions. You can pair an MCP server with Claude, GPT, or another supported model.

The Core MCP Components

• Host: The AI application that manages the model and MCP connections, such as Claude Code or an agent you build.
• Client: The component inside the host that communicates with one MCP server.
• Server: A local or remote process that exposes tools, resources, and prompts.
• Tools: Functions the model can call, such as searching an issue tracker, querying inventory, or creating a deployment.
• Resources: Readable context identified by URIs, such as documentation, files, or database-backed records.

MCP commonly uses JSON-RPC messages over local standard input/output or remote HTTP-based transports. Tool definitions include a name, human-readable description, and an input schema. That schema matters because it tells the host and model exactly which arguments are valid before a tool runs.

Hands-On Example: Expose a Support Lookup Tool

Suppose your team wants an AI coding assistant to retrieve customer support cases without giving it unrestricted database access. Create an MCP server that exposes only a get_ticket tool. The server can validate the ticket ID, call your internal API, and return only the fields needed for support work.

Tool shape:

name: get_ticket
description: Retrieve a support ticket by its numeric ID.
inputSchema: { type: "object", properties: { ticketId: { type: "integer" } }, required: ["ticketId"] }

Server-side handler logic:

if toolName is not "get_ticket", return an error
validate that ticketId is a positive integer
fetch https://support.example.internal/tickets/{ticketId}
return ticket subject, status, priority, and latest public comment

In a real Node.js MCP implementation, use an official MCP SDK to register the tool and connect it through the transport your host supports. Keep credentials in environment variables, enforce authorization in the server, and never rely on the model itself to decide whether a user may access a record. The MCP server is the security boundary.

Once configured in your host, a request such as “Find ticket 4821 and summarize the latest customer issue” can cause the host to list tools, select get_ticket, send { "ticketId": 4821 }, and use the returned data in its response. The model does not need direct database credentials or a custom prompt containing every ticket.

Using MCP with an Affordable Model API

Your MCP server can stay the same while you change the model behind the host or application. This is useful when tool use is frequent, because agent workflows can make several model calls while planning, calling tools, checking results, and producing an answer.

59API is a practical option for this architecture because it provides pay-as-you-go access to native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, plus GPT models. Its relay is compatible with Claude Code, Codex, and OpenAI SDK workflows, so developers can route model traffic through the API base URL https://api.59api.com while retaining their MCP tooling design.

For an OpenAI-compatible application, keep the MCP connection in your agent layer and configure the model client with your 59API key and the documented 59API base URL. Then select a model appropriate to the task: a fast lower-cost model for routine ticket summaries, or a stronger model for multi-step debugging and code changes. This separation makes it easier to control both permissions and spend.

Practical MCP Deployment Rules

• Expose narrow tools with clear input schemas instead of one unrestricted “run anything” tool.
• Return concise structured results so the model receives useful context without unnecessary tokens.
• Log tool calls, arguments, outcomes, and authorization decisions for debugging and auditing.
• Set timeouts and rate limits around external systems because tool calls can fail or be slow.
• Test ambiguous prompts to ensure the model cannot accidentally invoke destructive actions.

MCP gives AI applications a reusable way to work with real systems, while 59API helps keep the model side economical without sacrificing access to capable Claude and GPT options. Developers building tool-enabled agents can sign up for 59API when they are ready to connect their MCP workflows to low-cost, usage-based model access.

Preços · Docs