ar-agents is a Mercado Pago Agent Toolkit for the Vercel AI SDK 6. It ships 89 typed tools that an LLM agent can call directly to drive Mercado Pago billing flows: subscriptions, payments, refunds, checkout pro, marketplace OAuth, cuotas (installments), QR in-store, 3DS challenge resolution, point-of-sale devices, webhooks. Sidecar packages cover AFIP/ARCA, WhatsApp Business Cloud, banking (CBU/CVU + BCRA), and shipping (Andreani/OCA/Correo Argentino).

How is this different from the official mercadopago SDK?

The official mercadopago SDK is a thin REST client. It does not ship Vercel AI SDK tool schemas, does not implement webhook HMAC verification with replay protection, does not run on Edge Runtime (Node-only), and does not gate irreversible operations. ar-agents adds all of that on top of the underlying API: 89 typed tools, deterministic idempotency keys derived from inputs, programmatic human-in-the-loop on refund/cancel/delete, npm provenance attestation, Vercel KV adapters via subpath, OpenTelemetry instrumentation. You can use both packages in the same project; ar-agents wraps the underlying API directly without depending on the official SDK.

Does ar-agents work on Edge Runtime?

Yes. The whole package is Web Crypto-based with no node:crypto dependency, so it runs on Vercel Edge Functions, Cloudflare Workers, Deno, and any other V8-isolate runtime. Webhook signature verification, HMAC, and idempotency-key generation all use the Web Crypto API.

What is HITL (human-in-the-loop) in ar-agents?

Eight tools mutate state irreversibly (refund_payment, cancel_subscription, delete_customer_card, etc.). The toolkit accepts a requireConfirmation callback that gates each invocation: the tool function literally will not execute until your callback returns true. This is a programmatic gate, not just an LLM instruction. You can show the user a UI and wait for their explicit approval before any irreversible operation runs.

How does idempotency work?

Every POST request gets an auto-generated idempotency key. For LLM-driven retries, four mutating tools (create_payment, create_subscription, create_payment_preference, refund_payment) use a deterministic key derived from a SHA-256 hash of the meaningful inputs (external_reference, amount, payment_method, etc.). Same inputs produce the same key, so retries return the existing resource instead of double-charging the customer.

Yes. MIT license. No paid tier, no telemetry phone-home, no usage caps. The package is published to npm under the @ar-agents scope with SLSA v1 provenance attestations.

What about AFIP, WhatsApp, banking, shipping?

Sidecar packages cover the rest of the Argentine business stack: @ar-agents/identity (CUIT/CUIL validation + AFIP/ARCA padron lookup with monotributo category and IVA condition), @ar-agents/facturacion (AFIP/ARCA factura electronica via WSFE), @ar-agents/whatsapp (WhatsApp Business Cloud API with HMAC webhook verify and AR phone normalizer), @ar-agents/banking (CBU/CVU validation + BCRA Central de Deudores), @ar-agents/shipping (Andreani, OCA, Correo Argentino), @ar-agents/identity-attest (HMAC-signed verification orchestrator). Each ships independently to npm.

Is there a Model Context Protocol (MCP) server?

Yes. @ar-agents/mcp bundles all 7 ar-agents packages into a single MCP server compatible with Claude Desktop, Cursor, Codeium, Continue, Cline, or any MCP host. Auto-detects which packages to enable from environment variables. Listed on Glama (glama.ai/mcp/servers/ar-agents/ar-agents) and the official MCP Registry (io.github.ar-agents/mcp).

Una sociedad-IA en producción, demo ar-agents

2:30 · 6 paquetes · 0 humanos en el loop. Un agente Claude ejecutando el ciclo completo: constituir → CUIT → MP → factura → WhatsApp → BO. Datos mock, paquetes reales.

Qué muestra

El agente recibe una sola consigna: “Necesito constituir una sociedad-IA, conseguir CUIT, abrir cuenta MP, emitir factura y notificar al cliente por WhatsApp”. Sin más input humano, el agente atraviesa el ciclo llamando 6 paquetes ar-agents en secuencia:

@ar-agents/igj, verifica que el nombre “ACME-AI” está disponible en el dataset abierto de IGJ.
@ar-agents/identity, valida el CUIT y consulta el padrón AFIP/ARCA.
@ar-agents/mercadopago, crea el customer y la suscripción mensual de $50.000.
@ar-agents/facturacion, emite la factura electrónica con CAE.
@ar-agents/whatsapp, notifica al cliente con template de WhatsApp Business.
@ar-agents/boletin-oficial, suscribe el CUIT propio al firehose del BO.

Cómo está hecho

El video lo armé con Remotion: un framework de React para generar video programáticamente. El código está en apps/demo-video/ del monorepo: 6 scenes (intro, contexto, tesis, stats, terminal, outro) compuestas en una Composition de 1920×1080 a 30fps, total ~2:30. Las llamadas a tools en la pantalla del terminal son mocks deterministas, la lógica de cada paquete está en el monorepo para que cualquiera la verifique. Sin actores, sin foley, sin AI-generation: todo es código React renderizado frame-por-frame con Chromium headless.

Subtítulos y accesibilidad

El reproductor incluye subtítulos en español argentino (.vtt, activos por default). Para verlos sin sonido o en una pantalla muteada, el archivo bruto está en /video/sociedad-ia-demo.es.vtt: también CC0. Para traducirlo, abrí un PR contra apps/landing/public/video/sociedad-ia-demo.es.vtt.

Licencia

El video es CC0, dominio público. El código que lo genera es MIT. Republicalo, traducilo, subtitulá, citá libremente sin pedir permiso. La única restricción ética: no edites el video para tergiversar las capacidades técnicas mostradas.

Una sociedad-IA en producción.

Qué muestra

Cómo está hecho

Subtítulos y accesibilidad

Licencia