CatalystWatch: AI Agent That Explains Why Crypto Moves

The Problem With Crypto Alerts

Most crypto alert bots tell you what happened, price up 5%, volume spiked 3x and stop there. You're left opening five tabs trying to figure out why, usually too late to act on it anyway.

I built CatalystWatch to close that gap: an autonomous agent that doesn't just detect a market move, it investigates it.

What's Included

• Concurrent Scraping: Playwright pulls live stats for 10+ coins in parallel from CoinMarketCap, cutting cycle time significantly versus sequential requests.
• Anomaly Detection: A lightweight statistical layer flags real volume spikes (≥2x rolling average) and significant price moves (≥5% in 24h) with cooldown logic so the same move doesn't spam repeat alerts every cycle.
• News-Grounded Reasoning: Before the AI explains anything, it pulls real recent headlines for that coin (free, via Google News RSS, no API key). The agent is explicitly instructed to ground its "likely causes" in those headlines or admit it can't find one.
• Calibrated Confidence: This is the part I care about most. Most "AI explains the market" tools confidently invent a narrative regardless of whether the evidence supports it. CatalystWatch's prompt forces an honest answer: if a tiny price wobble has no real catalyst behind it, the agent says so and drops confidence to Low instead of overfitting a headline to noise.
• Self-Healing Pipeline: Discord delivery has exponential-backoff retry and rate-limit handling. If the Gemini reasoning step itself fails, the agent still ships a fallback alert with raw signals instead of silently dropping it.

How It Works

Scrape (Playwright) → Store (MongoDB) → Detect Anomaly → Fetch News → Reason (LangGraph + Gemini) → Alert (Discord)

Each stage is a node in a LangGraph state graph, so context (market stats, headlines, the alert itself) flows through cleanly and each step can fail independently without taking down the whole run.

A Real Example

When NEAR jumped 5%+ in a day, CatalystWatch didn't just say "NEAR is up." It pulled headlines about NEAR's AI-crypto narrative and a bullish price outlook published that week, connected them to the move, and rated its own confidence High because the evidence actually supported the size of the move.

Compare that to a flat, sub-1% wobble on another coin: same pipeline, but the agent correctly said no headline explained it and rated confidence Low. That distinction knowing when not to claim a cause is the whole point.

The Goal

I didn't want another bot that sounds confident regardless of whether it should be. I wanted an agent that reasons the way a careful analyst would: find the evidence first, size the conclusion to match the evidence, and say "I don't know" when that's the honest answer.

This project is for educational and research purposes only and does not constitute financial advice.