Prerequisites
Cryptohopper MCP configured in an MCP client — see the setup overview.
Explorer tier or higher for candle-based trend analysis per mentioned token. Ticker-only versions run on Pioneer — see subscription tiers.
A source of news — an article pasted manually, an RSS feed, a news API, or a Telegram/Discord channel you monitor.
Setup steps
Start with the manual-paste workflow
It's the fastest way to dial in the prompt before automating anything. Open your MCP client, paste an article, and use the prompt below.Here's a news article: [paste article here]
Using the Cryptohopper MCP, for each crypto token mentioned by ticker symbol:
1. Pull the current ticker from Binance.
2. Pull the last 100 × 4h candles.
3. Produce a one-paragraph take that answers:
- Is the token trending up, down, or ranging?
- Is 24h volume unusual vs. recent candle volume?
- Is the current price action consistent with what the
article is claiming?
At the end, tell me which (if any) of the mentioned tokens look like the market is actively reacting to this article, vs. already priced in, vs. ignoring it.
If a ticker is ambiguous (e.g. a project name that could refer to multiple tokens), flag it and ask rather than guessing.Review the first run
Check: did the agent identify all mentioned tokens, or miss some? Did it confuse any tickers? Is the market-reaction verdict defensible given the data?
Refine the prompt based on what failed
Common additions: "Note the article's publication timestamp and discount any reaction that happened before publication." / "If the article is a rehash of older news, say so." / "For tokens mentioned only in passing, skip them."
Automate
Three delivery modes in order of effort: manual paste (30 seconds per article); RSS-driven (scheduled script reads a feed and runs the prompt on new items); social stream (Twitter/X or Telegram, noisier, worth it for earliest signal). See how to schedule MCP workflows and how to send reports to Telegram, Discord, or email.
Filter output to reduce noise
Add to the prompt: "If no token shows meaningful market reaction, return a short 'no actionable signal' summary rather than writing a paragraph for each."
Cost profile
Per article processed (assuming 3–5 tokens mentioned):
Action | Explorer | Hero |
3–5 tickers | 3–5 units | 3–5 units |
3–5 × 100 4h candles | ~15–25 units | 3–5 units |
Per article | ~20–30 units | ~6–10 units |
20 articles/day × 5 weekdays | 2,000–3,000/week | 600–1,000/week |
Comfortable within Explorer's 30,000 calls/week. Pioneer cannot run the candle-based version but supports a ticker-only version at ~3–5 calls per article. See rate limits explained.
Troubleshooting
The agent identifies the wrong token
Long-tail tickers frequently collide. Add to the prompt: "If a ticker is ambiguous or you are not highly confident, skip it and mention it was excluded." Wrong identifications produce confident-looking but wrong analyses — the worst kind of failure.
Every article returns "market is reacting" verdicts
The agent is anchoring too hard on the article's framing. Add: "Apply a high bar for 'reacting' — require concrete numeric evidence (volume > 2× baseline, price move > 1 ATR). Default to 'no material reaction' unless the data clearly supports it."
The article is old but the agent treats it as new
Always pass the article's timestamp alongside the text, or have the agent extract it. Add: "If the article is older than 48 hours, focus on 'what happened since publication' rather than current state."
Output is too long to skim
Tighten per-token output to one sentence, not a paragraph. Only expand for tokens flagged as actively reacting. See the filter output step above.
Tokens are mentioned with project names, not tickers
Most models handle the mapping correctly, but if you see misses, add: "Also resolve any crypto project names you recognise to their most common ticker."
You want social-media input, not articles
Start with a small, curated set of accounts or channels and the same prompt. Budget for a higher false-positive rate and stricter filtering — social streams are richer but noisier than RSS feeds.
The scheduled run fires on every feed update, including trivial ones
Pre-filter on the feed side — skip items shorter than a threshold, skip items without any ticker-shaped token, skip items from low-signal sources. The MCP call is cheap but agent reasoning time is not free.