NAME
market-screener
Several times a day the screener fetches market data from different sources and checks it against fixed rules. Hits show up in a web interface and as a Telegram message, and through MCP I ask about them straight from an AI chat. The screener is an internal tool with no public access. I show it here because it holds almost everything an automation needs: schedules, several data sources, notifications, sign-in and an interface for AI.
FEATURES
- Scheduled runs
- Each rule set runs at its own time, before the market opens, after the close or in between. A scheduler starts the workers, and the results go into one shared database.
- Several data sources
- Prices, key figures, news and dates come from different services. Daily prices sit in a cache, so not every run loads everything again.
- Web interface with sign-in
- Hits as a table, details in a side panel. Several users, and each sees only the screeners shared with them.
- Telegram
- New hits arrive as a message on the phone.
- AI assessment
- On a click, a language model assesses a hit against the rules of its screener.
- MCP
- Read access for AI assistants. Ask Claude "Is X in a screener today, and why?" and it answers with the same figures the interface shows.
EXAMPLE
Sample output with made-up values. I do not show real hits here.
[16:10 ET] run: breakout · 412 symbols checked SYMBOL RULE DISTANCE STATUS ACME breakout +1.8 % triggered NORD momentum −0.4 % waiting KRAN gap +3.1 % triggered VOLT breakout −2.2 % dropped → 2 alerts sent to telegram · mcp: ready
DESIGN
- Rule sets as modules
- Each rule set is a package of its own and registers in one place. Neither the scheduler nor the interface knows the individual rule sets. A new one comes in without changing the rest, and a switch takes it out again.
- Tests for the structure
- CI checks the architecture too: no module may import another, and every setting the code reads has to be in the configuration.
- One database, three ways out
- The workers write to SQLite. A FastAPI reads from it for the interface, for Telegram and for MCP, so all three show the same figures.
- Decisions written down
- Every larger decision is an architecture decision record in the repository, more than 60 by now.
LIMITS
- The screener is an internal tool. There is no public access and no code to download.
- It gives no recommendations. It sorts data by fixed rules, and what follows from that is my call.
- It depends on third-party data sources. If one fails or changes its format, a run is missing until I adjust it.
STACK
python · fastapi · apscheduler · sqlite · vue · telegram · mcp