---

The problem

Dozens of AI / Agent projects blow up on GitHub every day, but you don't have time to read every README, scan every release, and decide what's actually worth a look.

This tool does it for you. On a schedule it scans GitHub Trending + Search, uses an LLM to filter for genuine "AI apps", does a medium-depth study of each new project, archives a structured report locally, and pushes the day's highlights to your phone. File-based storage, pluggable models, one-command self-hosting.

For: developers, technical teams, and research analysts who want to stay on top of the AI ecosystem without drowning in it.

✨ Features

• 🔍 Dual sources — GitHub Trending + Search API, merged and deduped, then LLM-filtered down to real "AI apps"
• 🧠 LLM deep-dive — reads each new project's README + key docs + releases and produces a structured analysis report
• 🗂️ Project library vs. daily — projects/ (per-project, date-independent, iterated over time) + daily/ (the day's highlights)
• 🔌 Pluggable models — Anthropic / OpenAI / Gemini / DeepSeek / OpenAI-compatible / Ollama; switch with one config line, with failover support
• ♻️ Smart revisits — when an old project trends again, it's only re-studied if there's a new release or a big star jump; otherwise just metadata is refreshed, saving tokens
• 📊 Live dashboard — a React + Tailwind board auto-published to GitHub Pages (Live Demo)
• 📤 Pluggable delivery — ships with a scheduled Feishu (Lark) bot DM; the channel is decoupled, so PRs for Telegram / Slack / Discord / email are welcome

🎬 Sample output

Each day produces a daily/<date>.md that's tidy and readable as-is:

# 🤖 AI Daily Digest · 2026-06-19

5 AI app projects today.

## 1. withastro/flue 🆕
> A complete TypeScript sandbox runtime for autonomous agents — not an SDK,
> but a next-gen agent architecture.

- 💡 Why look: solves tasks bare LLM calls can't, giving agents a sandbox +
  tools + skills + durable execution harness. For anyone building a
  Claude-Code-grade autonomous agent.
- 🏷️ agent-framework · sandbox-runtime · TypeScript · workflow-automation
- Language: TypeScript ｜ ⭐ 5714 (+305)
- Full report: projects/withastro__flue/analysis.md

Every project also accumulates a full analysis.md deep report under projects/<owner>__<repo>/ for when you want to go deeper.

🏗️ How it works

Design principle: collection and delivery are deterministic scripts (reliable, cheap); only the "study" step calls the LLM. Don't make the LLM do scraping or message-sending — it's expensive and flaky.

flowchart LR
    A[GitHub Trending<br/>+ Search API] --> B[Merge & dedupe<br/>data/index.json]
    B --> C{LLM filter:<br/>is it an AI app?}
    C -->|yes & new / big update| D[LLM deep analysis]
    C -->|old, no change| E[Update metadata only]
    D --> F[projects/&lt;repo&gt;/<br/>analysis.md]
    D --> G[daily/&lt;date&gt;.md]
    F --> H[React dashboard<br/>GitHub Pages]
    G --> I[Feishu / other<br/>scheduled delivery]

📁 Layout

projects/<owner__repo>/   # project library: metadata.json / analysis.md / quickstart.md / history.md / README.snapshot.md
daily/<date>.md|.json     # daily digest (json is consumed by the push step)
data/index.json           # global index + dedupe + revisit decisions
dashboard/                # React + Tailwind board, bundled into a single bundle.html
src/                      # Python pipeline
config/config.yaml        # config (focus scope, top_n, llm, delivery)
deploy/                   # launchd plist + run.sh

🚀 Quick start

# 1) Dependencies
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 2) Configure secrets
cp .env.example .env      # fill in GITHUB_TOKEN / your provider's LLM key / (optional) FEISHU_*

# 3) Run the full pipeline once (collect + study + daily report)
python -m src.pipeline

# 4) (optional) Push today's digest to Feishu
python -m src.push

# Debug a single stage
python -m src.collect          # just inspect what was collected

# Weekly trend report (aggregates historical metadata: hottest / longest-running / newcomers; --push to deliver)
python -m src.trend --days 7 --push

Works without Feishu: the pipeline still produces the daily/ digest and the dashboard — delivery is just an optional last step.

⚙️ Configuration (config/config.yaml)

• focus.search_topics / min_stars — focus scope; change here to expand categories later
• collect.top_n — max projects to deep-study per day (default 5)
• llm.provider / llm.model — switch models by editing just these two lines; keys live in .env
• analyze_revisit — when to re-study an old project

Required / optional secrets (.env)

In the Feishu console enable im:message and im:message:send_as_bot (add contact:user.base:readonly if you locate the user by email/mobile), then publish an app version.

🤖 Self-hosting / scheduling

Option A: launchd (macOS, an always-on local Mac)

# 1) Replace the PROJECT_DIR placeholder in the plists with this repo's absolute path
PROJECT_DIR="$(pwd)"
for f in pipeline push weekly; do
  sed "s#PROJECT_DIR#${PROJECT_DIR}#g" deploy/com.daily-news.$f.plist \
    > ~/Library/LaunchAgents/com.daily-news.$f.plist
done
chmod +x deploy/run.sh

# 2) Load
for f in pipeline push weekly; do
  launchctl load ~/Library/LaunchAgents/com.daily-news.$f.plist
done

• 06:30 runs the pipeline, 08:00 pushes (staggered for punctuality); Monday 08:30 sends the weekly trend
• ⚠️ Sleep delays triggers; keep the Mac awake / plugged in

Option B: GitHub Actions (no always-on machine — recommended)

The repo ships .github/workflows/daily-pipeline.yml: it runs the pipeline on a schedule, commits daily/ + projects/ back, and optionally pushes to Feishu. Three steps to enable:

1. Settings → Secrets → add GH_PAT and your provider's LLM key (e.g. ANTHROPIC_API_KEY); add FEISHU_* too if you want delivery
2. Settings → Actions → Workflow permissions → Read and write
3. Defaults to 06:30 Asia/Shanghai; cron is in UTC, so adjust for your timezone

You can also use a plain server cron calling python -m src.pipeline / python -m src.push.

🗺️ Roadmap

• M1 closed loop → M2 dual source + project library → M3 deep reports + daily digest → M4 automation + smart revisits
• GitHub Actions self-hosting / scheduling template (daily-pipeline.yml)
• More delivery channels (Telegram / Slack / Discord / email)
• Extensible to non-AI categories (config switch already reserved in the architecture)

Full design doc: docs/DESIGN.md.

🤝 Contributing

PRs and issues welcome! New delivery channels, data sources, and LLM providers are great entry points. Start with CONTRIBUTING.md (local setup + conventions) and SECURITY.md — this is a public repo; never commit any key or token.

📄 License

MIT © Steven