io.github.codespar/mcp-afip

The Problem

AI agents can write code, analyze data, and chat. But they can't operate a business — collect payments, issue invoices, ship products, or notify customers. Especially not in Brazil, where every service has its own API, auth pattern, and quirks.

Meanwhile, five categories of agentic payment infrastructure are shipping in parallel — checkout protocols, authorization layers, micropayment rails, identity frameworks, and issuing tools — and none of them compose cleanly.

MCP Dev Brasil bridges both gaps. Traditional Brazilian services + the new agentic payment protocols, all accessible through a single MCP interface.

The Solution

MCP Brasil gives AI agents typed tools to interact with Brazilian APIs and agentic payment protocols. Each server wraps a real service — payments, fiscal, logistics, messaging, banking, ERP, crypto, and now agentic protocols — so your agent can operate a complete business workflow.

🛒 Customer places order
  → 💳 Agent charges via Pix (Zoop)
  → 📄 Agent issues NFe (Nuvem Fiscal)
  → 📦 Agent generates shipping label (Melhor Envio)
  → 📱 Agent sends tracking via WhatsApp (Z-API)
  → 📊 Agent records in ERP (Omie)
  → 🏦 Agent reconciles balance (Stark Bank)

Six systems. Zero human intervention. One agent.

Agentic Payment Protocols

"The bridge looks more like middleware than a protocol." — The middleware is MCP.

Three new servers that bridge the emerging agentic payment stack:

The Autonomy Spectrum

Each protocol sits at a different level of agent autonomy:

 Human-in-loop ◄──────────────────────────────► Fully autonomous

  ┌─────────┐   ┌─────────┐   ┌─────────┐   ┌─────────┐
  │   ACP   │   │   UCP   │   │   AP2   │   │   x402  │
  │ Stripe  │   │ Google  │   │ Google  │   │Coinbase │
  └─────────┘   └─────────┘   └─────────┘   └─────────┘
  User confirms   Commerce     User sets      No user.
  every purchase  lifecycle    rules, agent   Machine-to-
  in-chat         managed      acts within    machine at
                  by agent     budget/scope   HTTP layer

The Convergence Stack

These protocols aren't competing — they're converging into layers:

┌─────────────────────────────────────────────┐
│  Application Layer        ACP / UCP         │  Chat UX, product discovery
├─────────────────────────────────────────────┤
│  Authorization Layer      AP2 / Mandates    │  Spend limits, audit trails
├─────────────────────────────────────────────┤
│  Tool Layer               MCP  ◄── WE ARE  │  Standardized agent tools
├─────────────────────────────────────────────┤
│  Settlement Layer         x402 / Pix / Card │  On-chain or traditional rails
└─────────────────────────────────────────────┘

CodeSpar sits at the Tool Layer — the middleware that connects every application, authorization, and settlement protocol through one interface.

Why this matters

Agent needs to buy something
  ├── Full commerce?       → Google UCP (search → cart → checkout → delivery)
  ├── Retail checkout?     → Stripe ACP (create_checkout → complete_checkout)
  ├── API micropayment?    → x402 (pay_request → USDC $0.001 → data returned)
  ├── Agent-to-agent?      → AP2 (authorize_payment → execute_payment)
  └── Brazilian merchant?  → Asaas / Zoop / PagSeguro (traditional rails)

All via MCP. Same interface. One agent.

Quick Start — Agentic Protocols

# Google UCP — full agentic commerce (early access)
npx @codespar/mcp-ucp

# Stripe ACP — agentic checkout (test mode, free)
npx @codespar/mcp-stripe-acp

# x402 — HTTP micropayments (testnet, free)
npx @codespar/mcp-x402

# AP2 — agent authorization (early access)
npx @codespar/mcp-ap2

Quick Start

With Claude Desktop

Add to ~/.config/claude/claude_desktop_config.json:

{
  "mcpServers": {
    "stripe-acp": {
      "command": "npx",
      "args": ["-y", "@codespar/mcp-stripe-acp"],
      "env": {
        "STRIPE_API_KEY": "sk_test_..."
      }
    },
    "zoop": {
      "command": "npx",
      "args": ["-y", "@codespar/mcp-zoop"],
      "env": {
        "ZOOP_API_KEY": "your-api-key",
        "ZOOP_MARKETPLACE_ID": "your-marketplace-id"
      }
    }
  }
}

With any MCP client

npx @codespar/mcp-stripe-acp   # Agentic Commerce Protocol
npx @codespar/mcp-x402         # HTTP micropayments
npx @codespar/mcp-ap2          # Agent authorization
npx @codespar/mcp-zoop         # Payments (marketplace, split)
npx @codespar/mcp-nuvem-fiscal # Fiscal
npx @codespar/mcp-melhor-envio # Logistics
npx @codespar/mcp-z-api        # WhatsApp
npx @codespar/mcp-omie         # ERP
npx @codespar/mcp-stark-bank   # Banking
npx @codespar/mcp-dev-brasil-api   # CEP, CNPJ (no key needed!)

Try it now (no API key)

BrasilAPI is free and public. Try it in your terminal:

npx @codespar/mcp-dev-brasil-api

Then ask your agent: "What is the address for CEP 01001-000?" or "Look up CNPJ 00.000.000/0001-91"

The Complete Loop

This is what makes MCP Brasil different — not individual connectors, but a complete business workflow across verticals:

To orchestrate all 6 steps with governance, approval workflows, and audit trails — use CodeSpar.

Servers

⚡ Agentic Protocols (4 servers)

💳 Payments (13 servers)

📄 Fiscal (3 servers)

📱 Communication (5 servers)

🇧🇷 Identity (1 server)

🏦 Banking (2 servers)

📦 E-commerce / Logistics (3 servers)

📊 ERP (3 servers)

🪙 Crypto / Stablecoins (5 servers)

🌎 LATAM-Wide (4 servers)

🇲🇽 Mexico (6 servers)

🇦🇷 Argentina (5 servers)

🇨🇴 Colombia (5 servers)

🔜 Coming Soon

Foxbit · BRLA · Coinbase · Transak · PagBrasil · Juno · NFe.io · PlugNotas · Movidesk · Infobip · Frenet · Loggi · Kangu · ReceitaWS · BigDataCorp · Sankhya · Totvs · Clip (MX) · Contpaqi (MX) · Ualá (AR) · Naranja X (AR) · Transbank (CL) · Nequi (CO expanded) · Rappi (CO)

Why MCP?

Model Context Protocol is the open standard for connecting AI agents to external tools. Instead of each agent building its own integrations, MCP provides a typed, discoverable interface that works with Claude, ChatGPT, Copilot, Cursor, and more.

AI Agent (Claude, ChatGPT, Cursor)
    ↕
MCP Server (this repo)
    ↕
Brazilian API / Agentic Protocol (Stripe ACP, x402, Zoop, etc.)

Each MCP server in this repo:

• Exposes typed tools with input/output schemas
• Handles authentication (OAuth, API keys, Basic Auth)
• Supports dual transport — stdio (default) and Streamable HTTP (--http flag)
• Compatible with Claude Managed Agents via MCP Connector
• Supports sandbox mode for safe testing
• Returns structured JSON responses

Running in HTTP mode

Any server can run as an HTTP server for remote/cloud use:

# stdio (default — local, Claude Desktop, Cursor)
npx @codespar/mcp-asaas

# HTTP (remote — Managed Agents, cloud deployments)
npx @codespar/mcp-asaas --http
# or
MCP_HTTP=true npx @codespar/mcp-asaas

HTTP mode exposes /mcp (Streamable HTTP) and /health (status check).

About CodeSpar

CodeSpar is an open source multi-agent platform that deploys autonomous AI coding agents to WhatsApp, Slack, Telegram, and Discord.

The MCP Generator in CodeSpar Enterprise can automatically generate MCP servers from API specifications — that's how this repo was bootstrapped.

Individual MCP servers are useful. Orchestrating many with governance is powerful. That's what CodeSpar does — including a Payment Gateway that integrates policy engine, payment routing, and mandate authorization across all rails.

Contributing

We welcome contributions! See CONTRIBUTING.md.

Want a server for a service not listed? Open an issue with the "server request" label.

License

MIT — use freely in commercial and open source projects.