Show & Tell: MCP Server for Brazilian Gov Procurement Data (stdio + MCPServerAdapter working example)

General
1

Hey CrewAI community!

Sharing a working example of the new MCPServerAdapter integration that might be useful for anyone building sales or prospecting agents.

The Problem:

Brazilian government procurement portals (where ~$500B in tenders are published annually) are notoriously hostile to scraping: malformed PDFs, expired SSL certs, weekly layout changes. Standard Playwright/Selenium approaches break constantly.

What I Built:

An MCP server that abstracts all the dirty work. It uses Google Dorks to find cached versions of procurement notices, processes them through Llama-3 for structured extraction, and serves strictly typed JSON to the agent.

The MCPServerAdapter Integration:

python
from crewai import Agent, Task, Crew
from crewai.tools.mcp_tool import MCPToolAdapter

# Connect to the MCP server via stdio
mcp = MCPToolAdapter(
    command="python",
    args=["mcp_server.py"]
)

sdr_agent = Agent(
    role="B2G Sales Hunter",
    goal="Find procurement opportunities in Brazilian cities",
    tools=mcp.get_tools(),  # Discovers tools automatically!
    verbose=True
)
```

**What the agent receives:**
```json
{
  "orgao_comprador": "Prefeitura de Campinas",
  "objeto_licitacao": "Contratacao de servicos de cloud computing",
  "valor_estimado": "R$ 850.000,00",
  "modalidade": "Pregao Eletronico"
}
```

The key architectural win is that the agent never touches a browser. All the messy extraction is encapsulated in the MCP server layer.

Full template with installation instructions on GitHub: https://github.com/guimaster97/crewai-gov-sales-agent

Happy to discuss the architecture or the MCP stdio integration if anyone is working on similar patterns!
1 Like