convez.ai
MCP ServerEscrita & ConteúdoAvançado
anbanai

Anban Writer Codex

anbanai/anbanwriter-codex

Toolkit IA para criação de conteúdo WeChat/Seednote com orquestração de agentes

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server para criação de conteúdo profissional em WeChat e Seednote com IA. Oferece 24 skills auto-descobertas e 6 subagentes Codex para pipelines completos de escrita, design visual e publicação multi-plataforma, integrando transcrição de áudio, geração de imagens e edição de vídeos.

Instalação

$npx convez-ai anbanai/anbanwriter-codex

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
50%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Criar artigos e posts para WeChat em escala
  • Gerar conteúdo viral otimizado para Seednote
  • Processar vídeos ao vivo, clips e edição automática
  • Gerar imagens de produtos para e-commerce
  • Otimizar SEO e análise de viralidade

Veredito

Solução robusta e completa para criadores de conteúdo em plataformas chinesas, com orquestração sofisticada de agentes e integração profunda com pipelines de mídia. Exige setup técnico considerável e conhecimento de Codex, mas entrega automação real de workflows complexos.

README

A tradução para o português deste README ainda está na fila.

Anban Creator for Codex

Professional WeChat and Seednote (种草笔记) content creation toolkit for OpenAI Codex — MCP server with AI-powered writing, visual design, and multi-platform publishing. This is the Codex-native port of claudecode/, connecting to the same anban-creator MCP server.

What you get

  • 24 auto-discovered skills (SKILL.md format): content writing, WeChat article assembly, Seednote viral analysis, live video slicing, video-use editing, line-art coloring, short-video cover replication, portrait pose variants, SEO, e-commerce product imagery, and more.
  • 6 native Codex subagents: end-to-end orchestrators (10,000+ word pipelines each) for the workflows above.
  • MCP integration: connects to the anban-creator HTTP MCP server for project management, image generation, WeChat publishing, TingWu/OpenAI-compatible FunASR transcription, and FFmpeg-driven clip assembly.
  • Lifecycle hooks: SubagentStop (per-agent delivery summary) + Stop (generic QA gate).

Prerequisites

  • Codex CLI installed (codex --version works)
  • ffmpeg + ffprobe (required for live-slicer, live-slice, and video-use)
  • jq (used by some skill-side validation steps)
  • An Anban Creator account at https://creator.anbanai.com — grab your API Key

Installation

1. Install the plugin

From a local clone:

codex plugin marketplace add ./codex
codex plugin install anban

2. Install the subagents

Codex plugins cannot bundle subagents (open limitation — see CODEX.md). The six subagents live in codex/agents/*.toml and must be copied to ~/.codex/agents/:

bash codex/install/install-subagents.sh

The script is idempotent and does three things:

  1. Copies agents/*.toml to ~/.codex/agents/, substituting __PLUGIN_ROOT__ with the discovered plugin install path (so each [[skills.config]] path resolves correctly).
  2. Merges install/agents-registration.toml into ~/.codex/config.toml — adds [features] multi_agent = true, [agents] max_threads = 6, and one [agents.<name>] block per subagent. Existing tables are preserved (no overwrites).
  3. Prints next-step reminders (env var setup, restart Codex, verify).

3. Set API credentials

Add to ~/.zshrc (or ~/.bashrc):

export ANBAN_API_KEY="<your-api-key-from-creator.anbanai.com>"
# Optional overrides:
# export ANBAN_API_URL="https://api.creator.anbanai.com"  # default
# export ANBAN_DEFAULT_PROJECT="<project-id>"             # skip list_projects

Then source ~/.zshrc (or restart your terminal).

4. Restart Codex

Quit and relaunch Codex so it picks up the new subagents, MCP server, and skills.

Verification

After restart, run:

/skills

Expected: Anban Creator skills listed (article, content-writing, seednote, ecommerce, video-use, line-art-coloring, etc.) with no "some skills omitted" warning.

/agents

Expected: 6 subagents listed (wechatarticle, seednote, designer, live-slicer, video, ecommerce) with their nicknames.

$anban-setup

Expected: the anban-setup skill runs list_projects and returns your configured projects. If it fails, your ANBAN_API_KEY is missing or invalid.

Usage

Implicit (skills auto-load)

Just ask in natural language:

写一篇关于 SwiftUI 与 Jetpack Compose 对比的文章

Codex's main agent detects the request, loads article, content-writing, topic-research, etc. on demand. No subagent is spawned.

Explicit (full pipeline via subagent)

For end-to-end runs that produce a complete publishable artifact, delegate to a subagent:

use the wechatarticle subagent to write a 3000-word article about Rust ownership
delegate to designer: colorize the line art at /path/to/lineart/ using a warm summer palette
use the live-slicer subagent on /path/to/live.mp4 — pull 5 high-density clips

The subagent runs autonomously, writes intermediate artifacts to $DIR/*.md, calls MCP tools, and emits a delivery summary on completion. You cannot interrupt mid-run (zero-interaction contract).

Hybrid

If you only want one capability (e.g. cover generation without writing the article), invoke the skill directly:

using the article-visual-design skill, generate a 2.35:1 cover for the article at ./draft.md

Capabilities by subagent

SubagentWhat it produces
wechatarticleResearched outline → final Markdown → WeChat-safe HTML → uploaded cover + content images → published draft
seednoteTopic/viral analysis → Markdown note (title + body + hashtags) → cover + 3-8 content images + tail image → archive
designerPer-lineart colored_NN.png + Color Bible + consistency report (PASS/MINOR/FAIL per entity) + manual-review flags
live-slicermetadata.json + audio.mp3 + cover.jpg + TingWu analysis + filtered sentences + clip plan + exported MP4s + CapCut drafts + transcript.md + summary.md
videoRoutes between dreamina-video generation, video-use footage editing, short-video-cover, portrait-pose-variants, and optional capcut-draft delivery
ecommerceProduct Bible (analyze product photos) → selling points (FABE) → asset plan → anchor-first generation with provider-adaptive ref strategy + vision self-check → compliance (广告法极限词) → archive + manifest

Troubleshooting

/agents shows nothing

  • Confirm ~/.codex/agents/*.toml exists (6 files).
  • Confirm ~/.codex/config.toml contains [agents.wechatarticle] etc.
  • Confirm [features] multi_agent = true is in config.toml.
  • Fully restart Codex (not just reload).

Subagent fails with "MCP tool not found"

  • echo $ANBAN_API_KEY — should print your key.
  • Confirm ~/.codex/config.toml has [mcp_servers.creator] (the install script does not add this; set it manually if you skipped plugin install):
    [mcp_servers.creator]
    url = "${ANBAN_API_URL:-https://api.creator.anbanai.com}/mcp"
    bearer_token_env_var = "ANBAN_API_KEY"
    

Skill paths in subagent TOMLs are wrong

The install script substitutes __PLUGIN_ROOT__ based on ~/.codex/plugins/cache/.... If you installed the plugin to a non-default location, set ANBAN_PLUGIN_ROOT and re-run:

ANBAN_PLUGIN_ROOT=/custom/path bash codex/install/install-subagents.sh

Hooks don't fire

Plugin-bundled hooks require explicit user approval in Codex. Check Codex's plugin trust UI; if hooks are still blocked, the same delivery summaries will run via the prompts embedded in each subagent's developer_instructions (so quality reports still happen — just less reliably).

Differences from the Claude Code plugin

See CODEX.md for the full mapping table. Summary:

  • Subagents only run when explicitly invoked (use the X subagent).
  • Subagents declare their own MCP servers and skills (no inheritance).
  • No TaskCompleted event — Stop matcher "*" replaces it.
  • Plugin hooks require explicit trust review.
  • Skills, themes, writers, MCP tool names, and content pipelines are byte-identical to claudecode/.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados