A 71,000-Star Stock Analyser That Runs on GitHub's Free Cron

Most "AI agent" projects are demos that need a human to press run. You clone the repo, install forty dependencies, and then discover that the intelligence is real and the operations are imaginary. Nobody wants to be the person who remembers to execute it every morning.
ZhuLinsen/daily_stock_analysis, which has gathered roughly 71,600 stars and is one of the more surprising entries on GitHub's weekly trending list this month, is the opposite kind of project. It is designed to be forgotten about. You fork it, fill in a few secrets, enable a workflow, and then a dashboard arrives in your chat app at 18:00 Beijing time on every trading day whether or not you remembered anything.
The system covers A-shares, Hong Kong, US, Japanese, Korean and Taiwan markets for a watchlist you specify. It pulls market data and news, runs an LLM over the combination, and pushes a decision dashboard to WeCom, Feishu, Telegram, Discord, Slack or email. MIT licensed, Python 3.10 and up, with a paper behind it on arXiv.
The setup is the product
The GitHub Actions path takes about five minutes and requires no local environment at all.
Fork the repo. Add repository secrets. STOCK_LIST is the only mandatory one, formatted as comma-separated codes across markets: 600519,hk00700,AAPL,7203.T,005930.KS,2330.TW. Pick at least one notification channel, WECHAT_WEBHOOK_URL, FEISHU_WEBHOOK_URL, TELEGRAM_BOT_TOKEN plus TELEGRAM_CHAT_ID, DISCORD_WEBHOOK_URL, SLACK_BOT_TOKEN plus SLACK_CHANNEL_ID, or EMAIL_SENDER plus EMAIL_PASSWORD. Optionally add a search key for news quality: SerpAPI is the recommended one, with Tavily, Bocha, Brave, MiniMax and a self-hosted SearXNG instance as alternatives.
Then go to the Actions tab, accept the prompt about enabling workflows, and run "每日股票分析" manually once to confirm it works. After that it fires on its own at 18:00 Beijing time on weekdays, skips non-trading days including A-share, Hong Kong and US holidays, and supports forced runs and resumable checkpoints for the edge cases.
The local path, for people who would rather own the process:
git clone https://github.com/ZhuLinsen/daily_stock_analysis.git && cd daily_stock_analysis
pip install -r requirements.txt
cp .env.example .env
python main.py
The CLI has a reasonable set of switches. --debug for verbose output, --dry-run to generate without pushing, --stocks to override the watchlist for one run, --market-review for the market-level summary rather than per-stock, --schedule for a long-running local scheduler, and --serve-only to run the web surface without analysis. There is a Docker image at zhulinsen/daily_stock_analysis for the middle ground.
Why the free data sources matter more than the model
The economics of this project are the interesting part.
Default market data comes from AkShare, Baostock and YFinance, all free. The README is explicit that log lines saying a source is unconfigured do not affect the run. Paid alternatives exist, TickPlus for A-shares and ETFs, Tushare for historical stability, Longbridge for Hong Kong and US fields like turnover and PE, but the system is designed to produce a usable dashboard without spending anything.
Compute is GitHub Actions minutes, which for a public fork running once per weekday is effectively free. News search is the only place where cost appears, and even there the free tiers of several providers cover a watchlist of a dozen tickers.
That combination (free data, free compute, free delivery) is why the star count climbed. The analysis is no better than a paid terminal. The marginal cost of running it is zero, so you can leave it on forever and see what happens.
What it actually outputs
The dashboard is a per-stock block with the LLM's read on price action, news sentiment, announcements and catalysts, plus a market-level review. The README is candid about the degradation rules: social sentiment and some news sources fall back or drop out when their APIs rate-limit, and the system logs the fallback rather than pretending the data was complete.
That honesty is rarer than it should be. Most AI dashboards present a confident summary regardless of whether the inputs arrived. This one treats missing sources as a state to be reported.
What it gets wrong
The coverage is broad on paper and uneven in practice, and the repo's own documentation is the best guide to where.
Market data quality depends on which feed answers. AkShare and Baostock are community-maintained and occasionally lag or change schema; YFinance covers US names well and thinner markets less so. Hong Kong and US fields like turnover ratio, volume ratio and PE need Longbridge credentials to be complete, which means an unconfigured setup will produce a dashboard with visible gaps for those markets.
The LLM step inherits whatever the news layer found. If your search keys are missing, the sentiment section degrades to price action and announcements, and a stock with heavy social discussion will read as quiet. The README is clear that missing sources are reported rather than silently smoothed, which is the right call, but it does mean two users of the same repo can get materially different dashboards for the same ticker on the same day.
And the analysis itself is a language model summarising public information. It is not running a valuation model, it does not have consensus estimates, and it cannot tell you whether a price move was an overreaction. What it can do is compress forty headlines, three announcements and a week of price action into a page you will actually read before the market opens.
The part worth copying
Strip out the stocks and you have a template for a whole category of products that is currently underserved: scheduled, private, single-purpose agents that push a result rather than waiting to be asked.
The components are trivial to describe. A cron. A handful of free data sources with graceful degradation. An LLM step that turns structured data into prose a human will actually read. A delivery channel the user already checks. A config file with no server to maintain.
Almost every "AI assistant" shipped this year is the opposite: a chat window that requires you to know what to ask. The cron-and-push shape inverts that. It assumes the user has better things to do than interrogate a model, and it earns its place by arriving at a fixed time with something specific.
The caveat that has to be stated
This is a hobby-grade analytical tool, not investment advice, and the repo says so. An LLM summarising public news about a stock is not a research analyst, and a dashboard that arrives every evening can create a false sense of rigour precisely because it is so regular.
Use it as a filter that tells you what changed. Anyone treating it as an oracle that says what to do is using it wrong, and the same goes for most of the AI products launched this month: the value sits in the reduction of reading, never in the conclusion.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.