AI-powered intelligence agent + content engine — turn daily noise into bilingual briefings, then auto-publish to your entire social media matrix.
AI is exploding. New models drop weekly. Crypto never sleeps. Your feed is 80% noise. You need signal — but reading everything yourself doesn't scale. And hiring an editorial team isn't an option.
IntelFlow runs every morning: an AI agent searches the web across the topics you define, writes a publication-quality briefing in your voice, and simultaneously publishes to WordPress, WeChat, Feishu, and Twitter — before you finish your coffee.
One API key. Plain-language setup. No scrapers to maintain. No subscriptions. No editorial team.
🧠 Intelligence system — daily briefings shaped by your analysis dimensions, in your editorial voice, so every issue sounds like you wrote it.
📡 Content engine / 自媒体矩阵 — bilingual articles (Chinese + English) auto-published across your entire platform stack. One pipeline feeds your blog, WeChat Official Account, Feishu knowledge base, and Twitter thread — simultaneously.
The only open-source agentic system that:
- Uses your LLM as an agent — it searches the web autonomously; no pre-built scrapers, no fixed RSS lists to maintain
- Generates native Chinese + English simultaneously — not translation; two independent editorial voices, one pipeline
- Onboards in plain language — type "AI tools, crypto, indie hacking" → AI agent suggests dimensions + sources, done in 60 seconds
- Publishes to 4 platforms in one command — WordPress, WeChat Official Account, Feishu, Twitter/X; no other open-source project does all four
- Supports 8 LLM providers with native web search — Claude, GPT-5, Gemini, Qwen3, Kimi, ERNIE, Zhipu, or local Ollama
- Self-hosted, production-grade, no vendor lock-in, no per-article fees
| Morning Brew / The Rundown | Feedly / Curated | hn-digest / newsletter-gpt | IntelFlow | |
|---|---|---|---|---|
| Bilingual native output | No — English only | No — English only | No — English only | Yes — Chinese + English |
| Multi-platform auto-publish | No | No | No | Yes — WordPress + WeChat + Feishu + Twitter |
| AI content engine / 自媒体矩阵 | Human editorial team | Not applicable | No publishing pipeline | Yes — one pipeline, all platforms |
| Custom analysis dimensions | No — fixed editorial | No — you curate manually | No — hardcoded topics | Yes — fully configurable |
| Authentic editorial voice | Human teams required | Not applicable | Generic AI output | Config file, no team needed |
| Agentic web search | No — manual curation | No — RSS only | No — static sources | Yes — LLM searches autonomously |
| Self-hosted + production grade | No | No | Hobby-grade, fragile JSON | Yes — failover built in |
| Cost | Subscription fees | Subscription fees | Free but limited | Free with Gemini/Ollama · Pay only your LLM API cost |
Commercial products like Morning Brew need large editorial teams for a consistent voice. Open-source alternatives are hobby projects with no publishing pipeline. IntelFlow is the only self-hosted system that delivers production-grade intelligence and distributes content across your entire 自媒体矩阵 — configured entirely through a file, no team required.
See IntelFlow running in production:
- English briefings: www.lizecheng.net
- Chinese briefings (Feishu): feishu.cn/wiki
Your output will look completely different — it's shaped by your dimensions, your AI model, and your configured editorial voice.
# 1. Clone
git clone https://github.com/lizecheng2021-maker/IntelFlow.git
cd IntelFlow
# 2. Install
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
# 3. Launch the Web UI — paste one API key, define your focus areas, done
python web/app.py
# Open http://localhost:5050That's the setup. To run your first briefing:
bash scripts/run_daily.shOne API key is all you need. Everything else — publishing targets, additional data sources, editorial persona — is optional configuration.
A full daily briefing run produces:
output/2026-03-11/
├── daily_en.md # English briefing (~4,000-5,000 words)
├── daily_zh.md # Chinese briefing (~4,000-5,000 words, independent editorial voice)
├── cover_en.png # AI-generated cover image
├── cover_zh.png # AI-generated cover image
└── briefing.json # Structured source data
Briefing structure (adapts to your configured dimensions):
30-Second Summary
─────────────────
[Your Dimension 1] e.g., AI Industry — 3-5 items, each with independent analysis
[Your Dimension 2] e.g., Crypto — signals, not summaries
[Your Dimension 3] e.g., SaaS — competitive moves, funding
...
Today's Synthesis Cross-dimensional insight, 400-600 words
Each item carries a judgment, not just a headline. The AI cross-references signals across dimensions and calls structural shifts when it sees them.
Output cadence:
| Format | Length | Cadence |
|---|---|---|
| Daily Briefing | 4,000-5,000 words | Every day |
| Weekly Deep-Dive | 8,000-10,000 words | Aggregated |
| Monthly Review | 12,000-15,000 words | Trend synthesis |
IntelFlow's core concept is dimensions — independent analysis tracks the AI uses to organize its research. Define them once in the Web UI.
Crypto Trader:
- Market Signals 30% | On-chain Data 25% | Regulatory 20% | DeFi Protocols 15% | Macro 10%
SaaS Founder:
- Competitor Intel 25% | Customer Pain Points 25% | Tech Stack 20% | Funding Landscape 15% | Growth Tactics 15%
Academic Researcher:
- Paper Releases 30% | Grant Funding 20% | Conference News 20% | Industry Applications 15% | Policy Impact 15%
Game Developer:
- Industry News 30% | Tech Releases 25% | Community Sentiment 20% | Competitor Moves 15% | Platform Changes 10%
Beyond dimensions, you can configure:
- Editorial voice — tone, style, recurring phrases, analysis depth
- Language output — Chinese only, English only, or both simultaneously
- Publishing targets — WordPress, WeChat, Feishu, or just local Markdown
- AI model — switch between Claude, GPT-5, Gemini, Qwen3, GLM-4.6, Kimi, or local Ollama with no code changes
IntelFlow Pipeline (~25 min)
============================================================================
COLLECT (parallel) PROCESS GENERATE (parallel) PUBLISH
______________________ ___________ ____________________ ________
| Web Search | | | | | | |
| RSS Feeds | | prepare | | Section 1 (EN+ZH) | | WP |
| Hacker News |-->| briefing |--+--->| Section 2 (EN+ZH) |-->| WeChat |
| GitHub Trending | | .py | | | Section 3 (EN+ZH) | | Feishu |
| Reddit | |___________| | | Section N (EN+ZH) | |________|
| YouTube Transcripts | | | |____________________|
| Custom Collectors | v | |
|______________________| AI WebSearch | v
verification | assemble_report.py
+------------+
Key design decisions:
- Each collector has a 10-minute timeout — one slow source never blocks the pipeline
- The AI model is pluggable — switch providers without changing any pipeline code
- Failed sections auto-retry once without affecting other sections
- Built-in collectors (RSS, HN, GitHub, Reddit, YouTube) work with no extra API keys
- Bilingual generation runs in parallel — not sequential translation
Extend with custom collectors:
# Create scripts/collect_mydata.py
# Accept --date and --output args, save output as raw_mydata.json
# That's it — the pipeline auto-discovers collect_*.py scripts| Provider | Recommended Models | Web Search | API Endpoint | Get Key |
|---|---|---|---|---|
| Anthropic | claude-opus-4-6 / claude-sonnet-4-6 / claude-haiku-4-5-20251001 | ✅ tool use | https://api.anthropic.com/v1/messages |
console.anthropic.com |
| OpenAI | gpt-5 / gpt-5-mini / o3 / o3-pro | ✅ MCP native | https://api.openai.com/v1/chat/completions |
platform.openai.com |
| gemini-2.5-pro / gemini-2.5-flash / gemini-2.5-flash-lite | ✅ google_search | https://generativelanguage.googleapis.com/v1beta |
aistudio.google.com | |
| Zhipu AI | glm-4.6 / glm-4.6v / glm-4.6v-flash | ✅ web_search plugin | https://open.bigmodel.cn/api/paas/v4/chat/completions |
bigmodel.cn |
| Alibaba | qwen3-max / qwen3.5-plus / qwen3-coder-next | ✅ enable_search | https://dashscope.aliyuncs.com/compatible-mode/v1 |
dashscope.aliyun.com |
| Moonshot | kimi-k2.5 / kimi-k2-0905-preview | ✅ $web_search | https://api.moonshot.ai/v1/chat/completions |
platform.moonshot.ai |
| Baidu ERNIE | ernie-5.0-thinking-preview / ernie-4.5 / ernie-x1.1-preview | ✅ baidu_search | https://aistudio.baidu.com/llm/lmapi/v3/chat/completions |
aistudio.baidu.com |
| Ollama | llama3.3 / qwen2.5 / deepseek-r1 | ❌ local only | http://localhost:11434/v1 |
No key needed |
GPT-4o was discontinued February 2026. Use
gpt-5-minias cost-efficient OpenAI option.
Estimated daily cost (one briefing, 4-5 dimensions):
| Choice | Daily cost |
|---|---|
| Ollama (local) | $0 — runs on your machine |
| Gemini 2.5 Flash / Flash-Lite | $0 — free tier covers typical usage |
| qwen3-max / glm-4.6 | ~¥0.5–2 |
| gpt-5-mini / Claude Haiku 4.5 | ~$0.10–0.30 |
| gpt-5 / Claude Sonnet 4.6 | ~$1–3 |
| Claude Opus 4.6 / o3-pro | ~$5–15 (not recommended for daily use) |
Start free with Gemini or Ollama. Upgrade the model only if you need deeper analysis.
Contributions welcome:
- New data collectors — More RSS feeds, APIs, or platform adapters
- AI model adapters — Add support for more LLM providers
- Publishing integrations — Substack, Medium, Ghost, LinkedIn, etc.
- Web UI improvements — Better setup flow, real-time progress
Please open an issue first to discuss significant changes.
| Use case | Cost |
|---|---|
| Personal use, self-hosted | Free |
| Learning, open-source contribution | Free |
| Commercial deployment (SaaS, agency, reselling) | $1,000 USD / deployment |
For commercial licensing: open an issue or contact via GitHub.
If IntelFlow helps you think better, consider giving it a star.
github.com/lizecheng2021-maker/IntelFlow