English | 日本語 | 中文

ANTEI — a personal AI creative studio

ANTEI is a multi-model creative workbench for images, video, audio and 3D. It puts the strongest generation models behind one interface, keeps every result in a permanent archive, and turns what you make into reusable assets — characters, styles, voices, prompts and LoRAs — that feed the next piece of work.

Live: www.anteisuba.com · English / 日本語 / 中文

---

Why ANTEI

• One studio, many models. Generate across image, video, audio and 3D model families from a single workspace, with each model's real parameters exposed instead of a lowest-common-denominator form.
• Creative control first. Reference images with explicit roles, per-model prompt dialects, a node canvas for multi-shot work, and LoRA recipes cloned faithfully from their source images.
• Nothing gets lost. Every generation is stored permanently with its prompt, model, parameters and lineage, then organised into folders, cards and recipes you can reuse.
• An assistant that works the tools. A built-in operator that discusses before acting, inspects your references, researches on the web, writes model-specific prompts and edits the workbench or canvas — with every change undoable and every paid generation confirmed by you.
• Your keys, your bill. Generation runs on your own provider keys (BYOK), encrypted at rest. ANTEI never silently switches a request onto a platform key.

---

Product tour

Studio

The day-to-day workbench for single pieces.

• Image — a natural-language workbench and a tag workbench (Danbooru-style, with live tag checking for NovelAI), side by side; reference images with identity / pose / style / content roles; aspect ratio and resolution tiers per model.
• Image editing — instruction edits, inpainting, object replacement, style transfer, text rendering, background removal, element extraction and upscaling.
• Video — text-to-video, first/last-frame and multi-reference modes, with per-model duration, resolution and reference limits enforced before anything is sent.
• Audio — text-to-speech with a reusable voice library, sound effects and music.
• 3D — single image to textured GLB, including a mesh-first preview path.

Node canvas — the director's desk

A canvas for long-form, multi-shot work: script → shot breakdown → per-shot image and video generation → edit desk.

• Four node types (text, image, video, audio) wired through typed ports, so references and scripts flow into the shots that need them.
• Script nodes project into shot nodes and re-project when the script changes.
• An edit desk with a timeline to assemble clips, trim, subtitle and render.
• One undo stack for everything — whether you or the assistant made the change.

LoRA workbench

Reproduce a look, then customise it.

• Browse and import LoRAs from Civitai and Hugging Face into a personal library.
• Clone a source image's recipe — base model, LoRA stack and weights, sampler, steps, CFG, hires pass — and generate the same image on our own runner.
• Runner base models include Anima (Base / Turbo), WAI-Illustrious-SDXL, Pony Diffusion V6 XL, SDXL 1.0, Z-Image Turbo and Krea 2 Turbo; each LoRA family gets prompts written in its own dialect.
• Runs on a ComfyUI runner hosted on Modal (shared monthly quota, no key required).

Assistant

A single operator engine shared by the image, LoRA, video and canvas workspaces.

• Discuss first, act when told. Questions get answers; instructions get changes; real conflicts get a short multiple-choice question.
• Five verbs: look (inspect references and results), research (web search with cited sources), ask, apply (edit prompts, models, specs, canvas nodes) and request generation — which always stops at a confirmation card, because only you can start a paid generation.
• Bring your own planner model: Claude (Opus 5.5 / Sonnet 5.5 / Fable 5.1), OpenAI GPT-6 family, Gemini 3.x Flash, DeepSeek and Grok.
• Round summaries and optional long-term memory carry decisions across turns without replaying the whole history.

Library

• Assets — a private library of everything you generated or uploaded, with nested folders, batch actions and in-place detail views.
• Cards — character, style and voice cards that keep identity and look consistent across generations and shots.
• Prompts — personal recipes with versions and lineage back to the works they produced.
• Gallery — optional public showcase; prompts are only shared when you publish a cleaned recipe.

Claude integration (MCP)

The canvas exposes a Model Context Protocol endpoint, so Claude can read a project, review clips and edit the timeline while you watch. MCP tools can never trigger paid generation.

---

Models

Model availability changes often; the source of truth is src/constants/models/ and the provider registry in src/services/providers/registry.ts.

Modality	Model families	Routes
Image	GPT Image 2 / 2.5, Gemini Nano Banana Pro / 2.1 / 2 Lite, FLUX.2 Pro / Flash, FLUX Kontext Max, Seedream 5.0 Pro / Lite, Ideogram 4.5, Recraft V4, NovelAI V4.5 / V5	OpenAI, Google, fal, Ideogram, NovelAI, Volcano Engine, BytePlus
Video	Seedance 2.0 / 2.5, Kling V3 / O3 (incl. video-to-video edit), Wan 3.0, HappyHorse, Gemini Omni Flash, MiniMax H3	fal, Google, Volcano Engine, BytePlus, MiniMax
Audio	Fish Audio S2 Pro, ElevenLabs Sound Effects v2, ElevenLabs Music v2	Fish Audio, ElevenLabs
3D	Rodin Gen-2.5, Hunyuan3D v3 / v3.1 Pro, TRELLIS 2, TripoSR	Hyper3D, fal
LoRA	Anima, Illustrious / Pony / SDXL, Z-Image Turbo, Krea 2 Turbo	ComfyUI runner on Modal

Where a model is offered both by its vendor and by a reseller, ANTEI prefers the vendor's own API.

---

Architecture

flowchart LR
  B[Browser<br/>Next.js App Router] -->|auth · validate · enqueue| A[Next.js on Vercel<br/>API routes + services]
  A --> DB[(PostgreSQL · Neon<br/>Prisma 7)]
  A -->|dispatch job| W[Execution Worker<br/>Cloudflare Workers]
  W -->|provider API| P[OpenAI · Google · fal ·<br/>Volcano · MiniMax · …]
  W -->|LoRA jobs| R[ComfyUI Runner<br/>Modal]
  W -->|store results| S[(Cloudflare R2)]
  W -->|authenticated callback| A
  A -->|assistant planner| L[LLM providers<br/>Claude · GPT · Gemini · DeepSeek · Grok]

• Worker-first execution. The web app only authenticates, validates, records the job and dispatches it. Long-running provider calls, polling and uploads run in a Cloudflare Worker that reports back through an authenticated callback, so nothing depends on a serverless function staying alive.
• Layered codebase. constants/ and types/ (Zod schemas) → services/ (the only layer that touches the database or external APIs) → hooks/ → components/. API routes do three things only: authenticate, validate, call a service.
• Server-side guarantees. Ownership checks, usage accounting and the money gate live on the server; the assistant and MCP tools are structurally unable to start a paid generation.

Layer	Technology
App	Next.js 16 (App Router, Turbopack), React 19, TypeScript
UI	Tailwind CSS 4, shadcn/ui, Motion, React Flow, Tiptap
Auth	Clerk
Data	PostgreSQL on Neon, Prisma 7
Storage	Cloudflare R2 (permanent archive, CDN-served)
Execution	Cloudflare Workers (generation, video render, image proxy), Modal (ComfyUI runner)
i18n	next-intl — English, Japanese, Chinese
Validation	Zod across API contracts, provider payloads and model output
Testing	Vitest, Testing Library, Playwright

---

Repository layout

src/
├── app/            routes (App Router) and API routes
├── components/     ui/ (stateless primitives) · business/ (stateful features)
├── constants/      models, providers, limits, routes — check here first
├── contexts/       studio and workbench state
├── hooks/          client-side state and data hooks
├── lib/            shared utilities and the API client
├── messages/       en / ja / zh translations
├── services/       server-only business logic and provider adapters
└── types/          Zod schemas and inferred types
workers/            Cloudflare Workers (execution, render-video, image proxy) and the runner
prisma/             schema and migrations
docs/               workflow, references and checklists (start at docs/README.md)

---

Local development

Requirements: Node.js 22, npm 10+, a PostgreSQL database (Neon recommended), a Clerk application and a Cloudflare R2 bucket.

npm install
cp .env.example .env.local   # fill in database, Clerk, R2 and encryption secrets
npm run dev                  # http://localhost:3000

Common checks:

npm run typecheck
npm run lint
npm run test:run

Provider keys are added per user in Settings → Keys inside the app; you only need platform keys in .env.local for the features that use them (for example the assistant's default Gemini route). The execution worker has its own package under workers/execution with separate tests.

---

Security and privacy

• Provider keys are encrypted with AES-256-GCM and only decrypted server-side for the request that uses them.
• Every API route authenticates with Clerk first and checks ownership on the server.
• Server-side fetches of user-supplied URLs go through an SSRF guard.
• Raw prompts stay private; public recipes are published explicitly and cleaned before they appear in the gallery.

---

Documentation

Engineering docs live in docs/: the task workflow, per-domain references (canvas, LoRA, assistant, providers, model catalogue) and release checklists.

License

This repository does not include an open-source license. All rights reserved.