See both sides run

Same infrastructure, same governance. Watch the direction of money flip while the mandate gate, the router, and the receipt stay identical.

↗ Your agent BUYS

↘ Your agent SELLS

✦ Either way, the same artifact

Every settled move on either side seals a signed receipt binding permission, settlement, and proof:

What is CodeSpar?

CodeSpar is the agentic operating system for money movement in Latin America. Commerce is the wedge; money movement is the platform. We provide a single SDK and a unified session model that lets your AI agents move money (commerce checkout, procurement, payroll, treasury and cross-border) and handle everything around it: process payments, issue invoices, ship packages, send notifications, and interact with ERPs, without integrating dozens of regional APIs yourself.

Think of CodeSpar as the Stripe for AI agent money movement. Where Stripe unified web payments behind one API, CodeSpar unifies the rails the region actually uses -- Pix, boleto, Nota Fiscal, Correios, Mercado Envios, WhatsApp Business, and more -- behind one SDK call that any AI agent can invoke.

import { CodeSpar } from "@codespar/sdk";

const codespar = new CodeSpar({ apiKey: process.env.CODESPAR_API_KEY });
const session = await codespar.create("user_123", {
  servers: ["stripe", "mercadopago", "correios"],
});

const tools = await session.tools();
// 24 tools ready for your agent to call

Who is CodeSpar for?

What can you build?

Buy-side — the agent spends:

Sell-side — the agent collects:

Architecture

CodeSpar sits between your AI agent and every major LatAm commerce API. Both faces converge here: a buy-side spend and a sell-side charge enter the SAME router, hit the SAME mandate and policy gates, and land in the SAME audit ledger — the side only changes which meta-tool the agent calls. The architecture has four layers:

Your Agent (Claude, GPT, Gemini, LangChain, CrewAI, or any MCP client)
    |
    v
@codespar/sdk  <-- Framework adapters: claude, openai, vercel, langchain, google-genai, mastra, crewai, autogen, llama-index, letta, camel, mcp
    |
    v
CodeSpar API (api.codespar.dev)
    |                                  Routing, auth, usage tracking, billing
    v
MCP Server Catalog
    |
    +-- Payments:    Zoop, Asaas, Pagar.me, PagSeguro, Cielo, Stone, EFI, iugu, Vindi, EBANX, Mercado Pago, Conekta, Wompi
    +-- Fiscal:      Nuvem Fiscal, Focus NF-e, Facturapi, AFIP, Siigo
    +-- Logistics:   Melhor Envio, Correios, Skydropx, Andreani, Coordinadora
    +-- Messaging:   Z-API, Evolution API, Zenvia, Take Blip, RD Station
    +-- Banking:     Stark Bank, Inter, Nubank, Nequi, BCRA, Pix BCB, Open Finance, STP/SPEI
    +-- ERP:         Omie, Bling, Tiny, Conta Azul, Colppy, Alegra, Bind ERP, Belvo
    +-- E-Commerce:  VTEX, Mercado Libre, Tienda Nube
    +-- Crypto:      Mercado Bitcoin, Bitso, Circle
    +-- Protocols:   Stripe ACP, Google UCP, x402, AP2 (protocol scaffolds)
    +-- Data:        BrasilAPI

How it flows:

1. Your agent receives a user request (e.g., "Create a R$49.90 checkout link for the Pro Plan").
2. The adapter converts the agent's tool call into a CodeSpar SDK call.
3. The SDK sends the request to api.codespar.dev, which authenticates, routes, and tracks usage.
4. The API forwards the call to the appropriate MCP server (e.g., the Stripe MCP server).
5. The MCP server executes the operation against the regional API and returns structured results.
6. Results flow back through the stack to your agent, which presents them to the user.

The Complete Loop

Every agent interaction follows a consistent three-phase pattern we call the Complete Loop:

const result = await session.execute("codespar_discover", {
  query: "charge a buyer in BRL via Pix",
});

{
  "matches": [
    { "tool": "codespar_charge", "server": "asaas", "score": 0.91, "connected": true },
    { "tool": "codespar_charge", "server": "mercadopago", "score": 0.88, "connected": true }
  ]
}

const result = await session.execute("codespar_charge", {
  method: "pix",
  amount: 4990,
  currency: "BRL",
  description: "Pro Plan",
});

{
  "payment_id": "pay_abc123",
  "pix_code": "00020126580014br.gov.bcb.pix...",
  "qr_code_url": "https://api.codespar.dev/qr/pay_abc123.png",
  "amount": 4990,
  "currency": "BRL",
  "status": "pending"
}

// Chain: send the Pix code via WhatsApp
const notification = await session.execute("codespar_notify", {
  channel: "whatsapp",
  to: "+5511999887766",
  template: "pix_charge",
  variables: {
    customer_name: "Maria",
    pix_code: result.pix_code,
    amount: "R$49.90",
  },
});

The 14 Meta-Tools

Meta-tools are high-level operations that abstract the underlying MCP servers. Instead of calling provider-specific endpoints, your agent calls a meta-tool and CodeSpar handles the routing. They split by the side your agent plays — the same grouping you see in the sidebar.

Your agent buys — the agent spends, under a signed mandate

Your agent sells — the agent collects, as the merchant

Shared rails — both sides, no money direction

Platform at a glance

• Every major LatAm commerce API covered as an MCP server — payments, fiscal, logistics, messaging, banking, ERP, e-commerce, crypto, public data, and agentic commerce protocols
• 14 meta-tools that abstract all server routing and provide a consistent interface across 60+ catalog rails
• Every major agent framework supported — Claude, OpenAI, Vercel AI SDK, LangChain, Google Gemini, Mastra, CrewAI, AutoGen, LlamaIndex, Letta, CAMEL-AI, plus raw MCP
• One SDK — @codespar/sdk on npm, with a single session model for all operations
• Published under @codespar — every server and adapter available as its own npm package
• Two environments — csk_test_ keys for sandbox with mock data, csk_live_ for production

Getting Started