Avcs

AI Visual Content Studio → Avcs

English · 简体中文

Summary

Avcs is a local-first visual content studio. It can use Codex Agent through codex app-server, or AvcsAgent through Vercel AI Gateway when Codex is not available.

Requirements

• Install Codex CLI to use the Codex Agent harness.
• Configure a Vercel AI Gateway API key in /web/settings to use the AvcsAgent harness.

Usage

Architecture

Avcs is a local-first web app. Elixir/Phoenix is the single local backend boundary for state, files, SQLite, and Agent access; the React frontend never reads the local filesystem, SQLite, codex app-server, or Vercel AI Gateway directly. Desktop packaging can use a Tauri shell with an ElixirKit-style bridge to start Phoenix and open the web UI.

┌────────────────────────── browser / Tauri shell ───────────────────────────┐
│ system tray + local-port web app in browser                                │
└──────────────────────────────────────┬─────────────────────────────────────┘
                                       │
                                       ▼
┌────────────────────────── Elixir/Phoenix + React ──────────────────────────┐
│ UI, WebSocket channels, HTTP APIs, app boundary                            │
└──────────────────────────┬───────────────────────────────────────┬─────────┘
                           │                Agent Harness          │
                           ▼                                       ▼
┌────────────────────────────────────────────────────┐  ┌────────────────────┐
│ SQLite                                             │  │ Codex Agent        │
└────────────┬───────────────────────────┬───────────┘  │ or AvcsAgent       │
             │                           │              └──────────┬─────────┘
             ▼                           ▼                         ▼
┌────────────────────────┐  ┌────────────────────────┐  ┌────────────────────┐
│ global DB              │  │ project DB + files     │  │ image generation   │
│ ~/.avcs/avcs.sqlite3   │  │ .avcs/project.sqlite3  │  │ tool events        │
│ settings + secrets     │  │ work/, output/         │  │ text streaming     │
└────────────────────────┘  └────────────────────────┘  └────────────────────┘

The /web/settings page controls the global harness mode: codex, auto, or avcs_agent. The default is codex. auto prefers Codex and falls back to AvcsAgent when Codex is unavailable and the gateway key is configured.

Codex Agent uses Codex built-in image_gen. AvcsAgent uses the Avcs backend image_gen tool, which writes generated files to the current project output/ directory and then records assets, chat items, and board items through Phoenix.

Why Avcs

Avcs started from three practical needs.

First, when I ran out of credits on Lovart, I realized I already had a ChatGPT Pro subscription. I did not want to pay for another dedicated image generation subscription on top of that, so Avcs is built around codex app-server as a local-first visual content studio.

Second, image generation often depends on accurate reference assets, not just prompts. If I want to create a poster for a specific Steam game, for example, I need the correct cover image available in the workflow. Avcs includes data provider support so projects can quickly load precise image assets from external sources and use them as references.

Third, local-first workflows can interact with the real contents of my local project folders. I can ask Avcs to read and understand the code in a project, then annotate a screenshot with button explanations or other inline help documentation that stays close to the project files it describes.

Demo

1. NASA APOD Poster

Avcs can load external visual references directly from the composer. Click the book-shaped icon in the lower-left corner, load the NASA APOD data provider, and ask Codex to generate a poster for the Astronomy Picture of the Day from a chosen month and day.

Open Data Providers	Load NASA APOD	Generate the Poster

2. Many-Case Output Board

Avcs keeps generated and imported visuals on a freeform Output board. A project can collect poster drafts, icon explorations, typography tests, annotated screenshots, and banner variants in one workspace, then select any image for preview or reuse as a reference in the next Agent turn.