GOAI 世界人工智能开源大赛 · 无界应用 Boundless Agents 赛道参赛作品
One product signal in → platform-native content out → real performance back → smarter next round.
一个产品信号进入 → 生成多平台原生内容 → 真实表现回流 → 驱动下一轮更聪明的生成。
中文说明见下方 / Chinese below · English first
Solo founders and lean teams ship product updates constantly — but turning one update into platform-native posts for RED (小红书), X, LinkedIn, Reddit, Instagram and more means rewriting the same story again and again, with no feedback loop from what actually performed.
Generic chatbots generate a post. They don't remember your brand, don't follow each platform's native format, don't track what happened after publishing, and never learn from it.
Finfold is an AI content workspace that closes the full growth loop for one-person and small growth teams:
product signal ──▶ agent workflow ──▶ 13 platform-native drafts
▲ │
│ ▼
private rules ◀── performance ◀── publish & track
- Workbench — enter one product update; get platform-native titles, body copy, CTAs and strategy notes for every selected channel, with brand & quality scoring.
- Brand Memory — brand positioning, audience, tone keywords, approved examples and banned phrases make output yours, and it compounds with use.
- AI Agent — a conversational agent (backed by Letta) that can decide what to create next, reuse brand context, and call product tools.
- Performance Loop — published content reports back; explicit user feedback and edit diffs feed the next generation round as private rules.
| Stage | How it works |
|---|---|
| Task understanding | A product update / growth goal is parsed into a creation brief |
| Orchestration | Agent workflow maps the brief to per-platform generation tasks |
| Knowledge augmentation | Brand Memory + Brand Rules are injected as grounding context |
| Tool calling | Generation, image, and content-iteration tools behind typed API routes |
| Result delivery | Editable drafts, quality scoring, export & publish-status tracking |
| Learning | User edits, ratings and (optional) performance data refine the next round — per-user, private |
- Framework: Next.js (App Router) + TypeScript + Tailwind CSS
- Agent / memory: Letta (agent memory & orchestration)
- Backend / DB / Auth: Supabase (Postgres, RLS, storage), deployed on Cloudflare (OpenNext)
- Payments (live product only): Creem checkout & webhooks
- Models: commercial LLM APIs via a provider-agnostic layer — see Open-source boundary below
This repository contains the competition release of Finfold:
- ✅ Included: application UI, agent workflow surface, platform adaptation layer, brand memory model, API route contracts, local-run setup, tests for included modules.
- 🔒 Not included (kept private): full production database migrations, internal analytics & founder dashboards, payment operations tooling, and unreleased features under active development.
- 💬 Commercial APIs are used for LLM generation and agent memory; all such calls sit behind provider-agnostic interfaces, with keys supplied via environment variables. No proprietary model weights are required to run the demo.
git clone https://github.com/joeymilano/Finfold.git
cd Finfold
npm install
cp .env.example .env.local # fill in your own keys (LLM API key is the only required one)
npm run dev # http://localhost:3000Live product (no setup): https://www.finfold.app
app/ # pages + API routes (workbench, agents, brand memory, generation)
components/ # UI: workbench, app shell, visual components
lib/ # platform rules, brand brain, content pipeline, agent data
hooks/ # client hooks (i18n, locale)
test/ # unit tests for included modules
docs/ # asset licenses & notes
- No scraping of private accounts; content generation is grounded in user-provided context.
- Per-user data isolation (row-level security in the live product).
- Guardrails: banned phrases, safety hints, and human review before anything is published.
Code in this repository is released under the MIT License. Third-party assets are listed in docs/asset-licenses.md.
独立开发者和小团队不断发布产品更新，但要把一次更新变成小红书、X、LinkedIn、Reddit、Instagram 等各平台原生的内容，意味着反复改写同一个故事——而且发布之后没有任何反馈闭环。
通用聊天机器人能生成"一篇"内容，但不记得你的品牌、不懂各平台原生格式、不追踪发布后的真实表现、更不会从中学习。
Finfold 是一个 AI 内容工作台，把"产品信号 → 多平台原生内容 → 发布追踪 → 表现回流 → 私有规则"压缩成一条低门槛增长闭环：
- Workbench 工作台：输入一次产品更新，一次性获得所选平台的原生标题、正文、CTA 与策略建议，附带品牌与质量评分。
- Brand Memory 品牌记忆：定位、受众、语气关键词、优秀样例、禁用词——用得越多，输出越像"你"。
- AI Agent：由 Letta 驱动的对话式智能体，能决定下一步创作什么、复用品牌上下文、调用产品工具。
- 表现回流：发布后的编辑差异与显性反馈进入私有规则，驱动下一轮生成。
- 赛事：GOAI 世界人工智能开源大赛 · 无界应用 Boundless Agents 赛道
- 细分场景：面向独立开发者与精简团队的跨境内容增长智能体（AI + 出海营销）
- 在线 Demo：https://www.finfold.app
git clone https://github.com/joeymilano/Finfold.git
cd Finfold
npm install
cp .env.example .env.local # 填入你自己的密钥（仅需一个 LLM API Key）
npm run dev- 不抓取私有账户数据；生成内容以用户提供的上下文为准。
- 线上产品按用户行级隔离（RLS）。
- 内置护栏：禁用词、安全提示，任何内容发布前均需人工确认。
本仓库代码以 MIT License 开源，第三方素材见 docs/asset-licenses.md。