Verificación y Auditoría Técnica · Grupo URBINA

Lo que se probó. Lo que no.

Cada contrato de proveedor lleva su nivel de evidencia y sus límites. Ningún endpoint de pago se ha presentado como verificado si no se ejecutó en este build. El catálogo normaliza documentación pública; no es una promesa de calidad.

v1.0snapshot 20 de octubre de 2026

Proveedores

contratos en el registro

Claims

afirmaciones con fuente

Verificado local

ejecutado en este repo

Sin probar

documentado o inferido

Alcance de este build. El modo demo usa FFmpeg local y no requiere claves. Los adaptadores de Muapi, KIE, Atlas, fal, Replicate y Kuaishou están implementados según documentación pública y snapshots estáticos, pero no se ejecutaron contra sus APIs. La publicación social tampoco se simuló: sin OAuth solo se genera un kit.

Escala de evidencia

Un peso mayor exige ejecución real, no una inferencia.

Verified localpeso 4

Exercised end-to-end against the real implementation in this repository.

Documentedpeso 3

Derived from public provider documentation plus a static client snapshot; not live-tested in this build.

Inferredpeso 2

Reconstructed from gateway naming, shared contracts, or reference-client patterns. Requires a live probe.

Unknownpeso 1

Not enough evidence. The adapter fails closed instead of guessing.

Proveedores y contratos

Endpoints, polling, mapeos y límites declarados por adaptador.

Local demo (FFmpeg)

local

Verified localsin clave1 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
LOCAL	ffmpeg gradients + drawtext	render demo plate

Polling

elapsed-time simulation · cada 4 s · límite 10 min

Mapeo de petición

prompt→ drawtext

aspectRatio→ canvas dimensions

batchSize→ image count

duration→ clip length

Mapeo de respuesta

idmock_<uuid>

statusqueued → processing → succeeded

outputs/api/media/mock/*

Evidencia

Produces real PNG and MP4 files without a provider key.
local-execution · src/server/mockMedia.ts + FFmpeg smoke test
Cached per prediction so repeated polling does not re-render.
source-inspection · src/generation/providers/mock.ts

Límites

Output is a visibly synthetic demo plate, not a model-generated frame.
Requires ffmpeg on PATH; falls back to SVG posters when unavailable.

Muapi

https://api.muapi.ai

Documentedrequiere clave2 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
POST	/api/v1/{model-endpoint}	submit
GET	/api/v1/predictions/{id}/result	poll

Polling

result endpoint · cada 4 s · límite 10 min

Mapeo de petición

prompt→ prompt

aspectRatio→ aspect_ratio

resolution→ resolution

duration→ duration

seed→ seed

negativePrompt→ negative_prompt

images→ image_url | images_list

endFrame→ end_image_url

audio→ audio_url

video→ video_url

batch→ num_images

Mapeo de respuesta

idrequest_id, id, prediction_id, data.request_id

statusstatus, state

outputsoutputs, output, result, data

Evidencia

Submit uses x-api-key and a model-specific path under /api/v1.
public-doc · public Muapi API documentation + reference client helper
Result responses reuse status/state and outputs/output keys.
source-inspection · src/generation/providers/muapi.ts

Límites

Not live-tested in this build.
Model endpoint slugs are catalog data and may drift from the provider.
Batch and reference-image semantics can differ per model.

KIE.ai

https://api.kie.ai

Documentedrequiere clave2 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
POST	/api/v1/jobs/createTask	submit
GET	/api/v1/jobs/recordInfo?taskId={id}	poll

Polling

recordInfo · cada 4 s · límite 10 min

Mapeo de petición

prompt→ input.prompt

aspectRatio→ input.aspect_ratio

resolution→ input.resolution

duration→ input.duration

images→ input.image_url | input.image_input

audio→ input.audio_url

video→ input.video_url

Mapeo de respuesta

iddata.taskId

statusdata.state, data.status

outputsdata.resultUrls, data.response.resultUrls, data.resultJson

Evidencia

Uses a job envelope with code/msg/data and a taskId.
public-doc · public KIE API documentation + reference client helper
resultJson is parsed as a secondary output channel.
source-inspection · src/generation/providers/kie.ts

Límites

Not live-tested in this build.
The resultJson shape varies by model.
Error code 200 is treated as success; other codes fail closed.

Atlas Cloud

https://api.atlascloud.ai

Documentedrequiere clave2 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
POST	/v1/predictions	submit
GET	/v1/predictions/{id}	poll

Polling

prediction endpoint · cada 4 s · límite 10 min

Mapeo de petición

prompt→ input.prompt

aspectRatio→ input.aspect_ratio

resolution→ input.resolution

duration→ input.duration

images→ input.image_urls

audio→ input.audio_url

video→ input.video_url

batch→ input.num_outputs

Mapeo de respuesta

idid, prediction_id

statusstatus, state

outputsoutput, outputs

Evidencia

Generic async prediction contract with Bearer auth.
public-doc · public Atlas Cloud API documentation + REA contract extraction
Payload uses a nested input object with image_urls.
source-inspection · src/generation/providers/atlas.ts

Límites

Not live-tested in this build.
Some catalog models may live on different Atlas routes; the adapter is intentionally generic.

Kuaishou direct (Kling)

env KUAISHOU_BASE_URL

Inferredrequiere clave2 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
POST	/v1/predictions	submit
GET	/v1/predictions/{id}	poll

Polling

prediction endpoint · cada 4 s · límite 10 min

Mapeo de petición

prompt→ input.prompt

images→ input.image_urls

duration→ input.duration

Mapeo de respuesta

idid, prediction_id

statusstatus, state

outputsoutput, outputs

Evidencia

There is no verified public Kuaishou direct contract in this build.
inference · adapter note
The adapter intentionally reuses the Atlas generic contract so it fails closed without a configured base URL.
source-inspection · src/generation/providers/atlas.ts

Límites

Do not use in production without a live contract probe.
A dedicated adapter is required if Kuaishou changes auth or payload shape.

fal.ai

https://queue.fal.run

Documentedrequiere clave3 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
POST	/{model-endpoint}	submit
GET	/{model-endpoint}/requests/{id}/status	poll status
GET	/{model-endpoint}/requests/{id}	fetch result

Polling

status then result · cada 4 s · límite 10 min

Mapeo de petición

prompt→ prompt

aspectRatio→ aspect_ratio

resolution→ resolution

duration→ duration

images→ image_url | image_urls

audio→ audio_url

video→ video_url

Mapeo de respuesta

idrequest_id

statusstatus

outputsresponse payload

Evidencia

Queue API uses Key auth and a request_id.
public-doc · public fal.ai queue documentation
Result URLs are constructed from the documented model path, never from an arbitrary response_url.
source-inspection · src/generation/providers/fal.ts

Límites

Not live-tested in this build.
Each fal model can define different input keys; the mapping is a lowest common denominator.

Replicate

https://api.replicate.com/v1

Documentedrequiere clave2 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
POST	/predictions (versioned) or /models/{owner}/{name}/predictions	submit
GET	/predictions/{id}	poll

Polling

prediction endpoint · cada 4 s · límite 10 min

Mapeo de petición

prompt→ input.prompt

aspectRatio→ input.aspect_ratio

resolution→ input.resolution

duration→ input.duration

images→ input.image | input.images

audio→ input.audio

video→ input.video

Mapeo de respuesta

idid

statusstatus

outputsoutput

Evidencia

Supports both official models and versioned community models.
public-doc · public Replicate HTTP API documentation
Submit sends Prefer: wait=5 to shorten the first poll.
source-inspection · src/generation/providers/replicate.ts

Límites

Not live-tested in this build.
Community model input schemas differ; the canonical mapping may not satisfy every model.

Higgsfield AI

https://api.higgsfield.ai

Documentedrequiere clave3 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
POST	/v1/generation	submit
GET	/requests/{id}/status	poll
POST	/requests/{id}/cancel	cancel

Polling

request status endpoint · cada 4 s · límite 10 min

Mapeo de petición

prompt→ prompt

aspectRatio→ aspect_ratio

resolution→ resolution

duration→ duration

batchSize→ batch_size / num_images

model→ model

startFrame→ start_frame

endFrame→ end_frame

characterReference→ character_reference

Mapeo de respuesta

idrequest_id, id, prediction_id

statusqueued / in_progress → completed | failed | nsfw

outputsimages[].url, video.url, videos[].url, output_urls

Evidencia

Official async generation API with request_id and status endpoint.
public-doc · docs.higgsfield.ai HTTP API reference
Standardized payload builder with media role conditioning.
source-inspection · src/generation/providers/higgsfield.ts

Límites

Not live-tested in this local build; requires HIGGSFIELD_API_KEY with active credits on console.higgsfield.ai.

Local GPU worker (WebSocket gateway)

env LOCAL_GATEWAY_URL (default http://127.0.0.1:8787)

Verified localsin clave2 endpoints · 3500 ms▾

Endpoints

Método	Ruta	Propósito
POST	/v1/jobs	submit
GET	/v1/jobs/{id}	poll

Polling

gateway job state · cada 4 s · límite 10 min

Mapeo de petición

canonicalRequest→ request

model→ model {id,endpoint,mode,kind,pipeline}

Mapeo de respuesta

idid

statusstatus

outputsoutputUrls

progressprogress

Evidencia

The gateway queues jobs and dispatches them over a WebSocket to connected workers.
local-execution · services/worker-gateway/server.mjs
Workers can return files as base64; the gateway persists them under /files.
local-execution · services/worker-gateway/server.mjs

Límites

The bundled worker renders a demo plate; replace its render() with ComfyUI or a real pipeline for production.
Transport is plain ws/http on a trusted LAN by default; terminate TLS at a reverse proxy for remote workers.

Artefactos analizados

Referencias de UI y metodología, con snapshot y límites.

Open-Higgsfield-AI (reference UI and model catalogue)

Autom8AI/Open-Higgsfield-AI

See upstream repository

snapshot 2026-10-20 static download

Static inspection of JSX components, CSS tokens, model arrays, and API client helpers. No runtime instrumentation.

Observaciones

Single prompt bar switches between text-to-image and image-to-image based on attached media.
Cinema Studio exposes camera body, lens, focal length, aperture, lighting and movement as separate controls.
The reference visual system uses #050505 background, #ff2d2d primary, glass panels and uppercase micro-labels.
The model list is a flat ESM array with per-mode slices (t2i, i2i, t2v, i2v, lipsync, v2v).

Límites

The snapshot is a point-in-time copy. Upstream endpoints and model lists can change without notice.
No provider traffic was captured; model endpoints were normalised rather than validated live.

Artefactos (7)

packages/studio/src/components/ImageStudio.jsx
packages/studio/src/components/VideoStudio.jsx
packages/studio/src/components/CinemaStudio.jsx
packages/studio/src/components/LipSyncStudio.jsx
packages/studio/src/models.js
src/styles/global.css
src/lib/api.js

REA: Reverse Engineer Anything (methodology)

morluto/rea

MIT

snapshot 2026-10-20 static download

Read AGENTS.md, README.md and tool contracts to adopt evidence grading, coverage limits, and fail-closed behaviour.

Observaciones

Every conclusion must carry evidence and an explicit statement of coverage and limitations.
Distinguish observation, derivation, inference and unknown; never present a simulation as real-provider verification.
Analysis should be read-only unless the user grants mutation authority.
Fail closed when authority, scope or identity is unresolved.

Límites

REA was used as a methodology reference, not as a runtime dependency of this app.
No native binary or Electron target from Higgsfield was analysed in this build.

Artefactos (5)

README.md
AGENTS.md
docs/mcp-contracts.md
docs/tool-design.md
docs/testing.md

Verificación reproducible

Estos comandos son los que respaldan el nivel «verified-local».

Comandos

Type safety

npm run typecheck

Unit and contract tests

npm test

Production build

npm run build

Demo media pipeline

npm run mock:media

Local worker gateway

npm run worker:gateway

Verificado en este build

Catalog compilation and mode resolution
Cost estimation in credits and USD
Optical prompt compilation
Credit reserve → settle and reserve → refund lifecycles
Mock image and video rendering through FFmpeg
Uploads, range requests, jobs, elements, projects and agent planning
Local worker gateway job round-trip (demo renderer)

No verificado

Any paid provider endpoint (Muapi, KIE, Atlas, fal, Replicate, Kuaishou)
Real OAuth publishing to TikTok or Instagram
GPU inference throughput or model quality
Cross-instance concurrency (the JSON store is single-process by design)

This file records how the platform's provider contracts were derived. No paid provider API was called while building this demo. Contracts marked 'documented' come from public API documentation and the reference client snapshot; they have not been live-tested here. Contracts marked 'inferred' are best-effort reconstructions and must be validated before production use. 'verified-local' contracts were exercised end-to-end in this repository.

Conectores y entrega ·Volver al estudio