AFIP / ARCA Constancia de Inscripción

Drive the public ARCA Constancia de Inscripción form and return one JSON envelope: the parsed constancia plus the official PDF. No Clave Fiscal — this is the form any citizen can use with just a CUIT.

Required: BROWSERBASE_API_KEY env var and browse CLI installed.

Typed companion: the same lookup is available as a typed, testable npm package — @ar-agents/constancia, part of the Arg toolkit — for agents on the Vercel AI SDK. Both share the Output contract below (parseSkillOutput). One artifact, two surfaces: use the package for programmatic/Vercel-AI use, this skill for any browser-driving agent.

When NOT to use this skill: if the caller only needs the data (not the PDF) and has an AFIP X.509 cert provisioned, the SOAP padrón webservice (@ar-agents/identity lookup_cuit_afip) is faster and needs no browser. This skill's edge is the PDF artifact and the no-cert path.

Output contract (STRICT — the companion package parses this verbatim)

Print exactly one JSON object to stdout as the last line, nothing after it. Shape:

{
  "cuit": "20123456786",
  "found": true,
  "denominacion": "APELLIDO NOMBRE | RAZÓN SOCIAL",
  "tipoPersona": "fisica | juridica",
  "condicion": "Monotributo | Responsable Inscripto | Exento | No Alcanzado | Sin inscripción",
  "monotributoCategoria": "A",
  "domicilioFiscal": { "direccion": "", "localidad": "", "provincia": "", "codigoPostal": "" },
  "actividades": [{ "codigo": "620100", "descripcion": "", "principal": true }],
  "impuestos": [{ "descripcion": "MONOTRIBUTO", "desde": "2026-04-17" }],
  "fechaInscripcion": "2026-04-17",
  "estado": "ACTIVO",
  "pdf": { "base64": "JVBERi0x…", "url": "", "codigoVerificador": "" }
}

Rules:

• CUIT not registered → { "cuit": "...", "found": false } (nothing else

required). Do not invent a constancia.

• Any failure (navigation, captcha, ARCA down) → `{ "cuit": "...", "error":

"<short reason>" }. If the page literally says the CUIT no figura inscripto, put that text in error — the parser maps "no figura / not found / inexistente" to a clean cuit_not_found`.

• condicion is free text — emit what the page shows; the parser normalizes it.
• pdf.base64 is preferred (durable). pdf.url only if you cannot capture

bytes (ARCA PDF URLs are short-lived). Include codigoVerificador when the document prints one.

• Never fabricate a field. Omit unknowns rather than guessing.

Procedure

Step 0 — Normalize and sanity-check the CUIT

CUIT_RAW="$1"
CUIT="$(printf '%s' "$CUIT_RAW" | tr -cd '0-9')"

If CUIT is not exactly 11 digits, emit {"cuit":"<raw>","error":"invalid_cuit: not 11 digits"} and stop. Do not open a browser for a malformed CUIT — browser runs are expensive and ARCA throttles.

Step 1 — Open the public form

Start a cloud session and open the public entry point (no auth):

browse open "https://www.afip.gob.ar/genericos/constanciainscripcion/" --cloud

Do not hardcode internal POST endpoints — ARCA rotates them. Navigate via the on-page elements. Locate the control that begins the consulta (a "Consultar" / "Ingresar" button or a CUIT field) semantically, not by a brittle fixed selector:

browse snapshot          # inspect the live DOM

Step 2 — Submit the CUIT

Type the 11-digit CUIT into the CUIT/identification field and submit. If a captcha / "no soy un robot" appears, rely on the Browserbase verified session + CAPTCHA solving (see the browser skill); do not attempt to bypass it manually. If it still blocks, emit {"cuit":"<cuit>","error":"captcha_blocked"} and stop.

Step 3 — Detect "no figura inscripto"

If the result page states the CUIT is not registered / "no figura inscripto" / "inexistente", emit {"cuit":"<cuit>","found":false} and stop. This is a normal, expected outcome — not an error.

Step 4 — Extract the constancia

From the rendered constancia, read: denominación, persona física vs jurídica, condición/régimen, monotributo categoría (if shown), domicilio fiscal, actividades (código + descripción + which is principal), impuestos (+ fecha desde), fecha de inscripción, estado, and the código verificador. Dump the raw page text/DOM to a temp file — do not hand-parse in a fragile pipeline:

browse get markdown > /tmp/constancia_${CUIT}.txt

Step 5 — Capture the PDF

Trigger the constancia's print/PDF action and capture the document as bytes:

browse pdf --output /tmp/constancia_${CUIT}.pdf
PDF_B64="$(base64 -i /tmp/constancia_${CUIT}.pdf | tr -d '\n')"

If byte capture is impossible, fall back to the visible PDF URL. A constancia without its PDF is still useful — never discard the parsed data because the PDF step failed; just omit pdf or set error only if nothing was obtained.

Step 6 — Emit the envelope

Build the JSON per the Output contract from the extracted fields + the base64 PDF, and print it as the final stdout line via the helper (it enforces the exact shape the companion package expects):

node "$(dirname "$0")/scripts/emit_constancia.mjs" /tmp/constancia_fields_${CUIT}.json

Gotchas

• ARCA = ex-AFIP (renamed 2025). The public form still lives under

afip.gob.ar. Treat the names as interchangeable.

• ARCA changes the form layout without notice. Navigate semantically; if the

page is unrecognizable, emit error: "fetcher_unexpected_response: <detail>" rather than guessing — the companion package treats wrong data as worse than no data.

• ARCA rate-limits. Do one attempt per CUIT per run. Never loop.
• Client-side scraping is CORS-blocked and trips anti-bot. Always run the

browser server-side / in Browserbase (this skill does, via --cloud).

• The constancia is point-in-time. Emit what the form shows today; the caller

decides freshness policy.