🔌 Your first MCP bridge

In module 2.1, Claude called a script. It works, but the agent needs to know the file name and construct the command. With MCP it’s different: the agent receives a list of ready-to-use tools, each with a name, description, and fields.

R3 combines two steps on the ladder. Reading the file is a local bridge (step 6). The way to deliver it to the agent is MCP (step 2). The result: a system without an API starts to look, to the agent, like a system with official tools.

How to read the diagram: on the left, the person making the request; in the middle, the bridge, which knows the rules; on the right, the files. The dashed arrows to the files are read-only: the bridge never writes to them.

The model lives in runtime/pontes/mcp-modelo/server.mjs. It's an MCP server with no dependencies: you don't need to install any packages, just have Node.

It includes two example tools, one for each course character. The description of each tool is what the agent reads to decide when to use it.

What to look for in the table: the fields are optional. Without data, the first lists all available times; without agrupar_por, the second groups by client.

data,hora,profissional,status,paciente
2026-10-06,08:00,Dra. Ana,ocupado,Paciente 01
2026-10-06,09:00,Dra. Ana,livre,
2026-10-06,10:00,Dra. Ana,livre,
2026-10-06,14:00,Dr. Bruno,ocupado,Paciente 02
2026-10-06,15:00,Dr. Bruno,livre,
2026-10-07,08:00,Dra. Ana,livre,
2026-10-07,09:00,Dra. Ana,ocupado,Paciente 03
2026-10-07,16:00,Dr. Bruno,livre,

Same rule as module 2.1: start without an agent. The server has a self-test mode, --selftest. It lists the tools, calls each one with sample data, and prints the result.

It doesn't use any quota: no AI model is involved in this test. It's just Node reading the CSVs.

node runtime/pontes/mcp-modelo/server.mjs --selftest

tools: 2 (listar_horarios_livres, resumo_vendas)
2026-10-06 09:00 · Dra. Ana
2026-10-06 10:00 · Dra. Ana
2026-10-06 15:00 · Dr. Bruno
TOTAL: R$ 856.00

The three times match the lines livre from 10/06 that you saw in the previous section. The total for Sônia still needs to be checked. The self-test shows only the last line of the summary; the table below recalculates it manually, row by row from the erp-vendas.csv.

What to look for in the table: the hand calculation reaches the same TOTAL: R$ 856.00 from the bridge (the period instead of a comma is just the program's format).

In the kit, this part is already set up. Two files handle it: the .mcp.json, at the root, records the bridge; the .claude/settings.json allows its use.

When you open claude in the folder, it reads the .mcp.json and starts the server on its own. You don't need to leave anything running.

{
  "mcpServers": {
    "ponte-modelo": {
      "command": "node",
      "args": ["runtime/pontes/mcp-modelo/server.mjs"]
    }
  }
}

"enabledMcpjsonServers": [
  "ponte-modelo"
]

claude mcp list

ponte-modelo: node runtime/pontes/mcp-modelo/server.mjs - ✔ Connected

Now the agent comes in. The command below uses claude -p: sends a single request without opening the screen and prints the response. It asks for the available times on 07/10, a day different from the self-test.

Changing the day is intentional. If the agent gives the right answer for a day you haven’t tested, that’s a sign it called the tool instead of repeating something it had already seen.

claude -p "Use a tool listar_horarios_livres da ponte-modelo e diga os horários livres de 2026-10-07."

How to read the diagram: the glowing box is the only point where the AI makes a decision: which tool to call and with what date. The filter in step 3 is code, not a guess.

Sônia does the same with the other tool. With claude open in the kit folder, it prompts: "Use the resumo_vendas tool from ponte-modelo to group by customer and tell me who bought the most." The answer has to match the table in topic 3: Mercado Sol is in the lead, with R$ 393,00.

Codex doesn't read the .mcp.json. It has its own log, created by a command. The bridge is the same: no line from the server.mjs changes.

Take a look at "$PWD/…": the recipe registers the server's full path in Codex, starting from the root of the disk, not the short path that the .mcp.json uses. That way, the record doesn’t depend on where Codex was opened.

codex mcp add ponte-modelo -- node "$PWD/runtime/pontes/mcp-modelo/server.mjs"

What to look for in the table: only the way you register it changes. That’s why MCP is worthwhile: you write the bridge once, and any agent that speaks MCP can use it.