No, AI can’t predict crypto from news. What this crypto sentiment analysis tool does is read every fresh headline about the coins I hold, score the vibe with Claude Haiku, and write the result to an Airtable so I have a running pulse instead of 47 open browser tabs. It runs five times a day on my n8n instance, costs me about $4 a month, and gives me peripheral awareness without making me sit through a single CoinTelegraph video.
I started this because I hold a handful of coins and I was spending way too many evenings catching up on what each one had been doing all day. Reading every coin in my portfolio top to bottom every night would eat a couple hours. I don’t have a couple hours for that. I have a million ideas and a small portfolio and a strong preference for staying informed without making it a part-time job.
So I built Crypto Alpha Pulse. It pulls 24 hours of news from three RSS feeds, grabs current prices from CoinGecko, asks Claude Haiku to score the sentiment per coin and across the macro picture, and dumps the scores into Airtable where my Crypto Alpha Dashboard reads from. Alpha the wolf (the mascot, because of course there’s a mascot) shows me the verdicts in one glance.
What this crypto sentiment analysis tool does
In plain order, every time the workflow runs:
- Pulls my holdings from Airtable so it only scores coins I own
- Reads the last 24h of crypto news from CoinTelegraph, Decrypt, and BeInCrypto
- Fetches live USD prices and 24h change for every held coin from CoinGecko (free tier, no key)
- Filters the news to per-coin buckets (a Bitcoin headline goes into the BTC bucket, not into Ethereum’s)
- Sends each coin’s bucket to Claude Haiku 4.5 with a structured prompt
- Parses the JSON response into a score (-100 to +100), a verdict (Strong Bull through Strong Bear), and a one-sentence action recommendation
- Sends the aggregated picture plus the macro/regulatory news to Claude one more time for an overall portfolio vibe
- Writes everything to three Airtable tables: per-coin snapshots, the macro pulse, and a deduped news log
21 nodes total. The whole run finishes in about 90 seconds.
Can AI predict crypto sentiment from news?
It can’t predict the price. Anyone telling you their AI predicts crypto is selling you something. What AI is genuinely good at is reading a pile of news in seconds, summarizing the vibe, and flagging the headline you would have missed if you scrolled past it on lunch break. That’s what this workflow does. It surfaces, it doesn’t forecast.
The score Claude returns is a sentiment read on the news, not a buy signal. “Bull” means the news is leaning positive in the last 24 hours. It does not mean buy. It means “the room is leaning bullish right now, do with that what you will.” I keep that distinction sharp because the second you start treating a sentiment score as a trade trigger, you’ve stopped using a peripheral awareness tool and started LARPing as a quant fund.
What this workflow is NOT
Worth saying upfront because the angle is honesty:
- It is not a trading signal. It scores news sentiment. News sentiment is a piece of context, not a decision.
- It does not predict price. Claude is reading words, not order books.
- It is not financial advice. I built this for me. I’m a builder who happens to hold some crypto, not a financial advisor.
- It will not save you from a flash crash. RSS feeds lag breaking news by minutes to hours.
What it IS: a way for someone who holds a few coins to keep peripheral awareness without burning evening hours on long reads. Breadth, not depth. That’s the trade I wanted.
n8n + Claude Haiku vs other approaches
Paid sentiment services exist and they’re built for active traders, not casual holders. Side-by-side:
| Option | Cost | What you get | Best for |
|---|---|---|---|
| Paid sentiment service (Santiment, LunarCrush Pro) | $30-100/month | Polished dashboards, social signals, historical charts | Active traders who watch all day |
| This n8n workflow + Claude Haiku | ~$4/month | Per-coin sentiment for YOUR holdings, runs 5x/day, your own Airtable | Casual holders who want a daily pulse |
| Manual reading (CoinTelegraph + Twitter) | $0 + ~60 min/day | Whatever you can absorb in an hour | People with way more time than me |
| Reading nothing | $0 + 0 min | A surprise every time you open Coinbase | Nobody’s real strategy |
The middle row is the spot I wanted: cheap, scoped to MY portfolio, queryable in Airtable later.
What it costs to run
Real numbers from my setup:
- Claude Haiku 4.5 API: ~$0.025 per run (5 per-coin calls + 1 macro call, ~12k input + 3k output tokens)
- 5 runs per day: ~$0.135/day, ~$4/month
- CoinGecko, RSS feeds, Airtable free tier: $0
- n8n instance: I’m already self-hosting for ~$8/month, this rides on top
Total marginal cost: about $4 a month. Less than one coffee. If you’re not on n8n yet, the n8n cost calculator shows whether self-hosting or cloud wins for you.
Why 5 runs a day, not hourly
I called the workflow “Pulse” because hourly was the original plan. In practice, running every hour was overkill. Crypto news doesn’t drop that fast, and rerunning the same 24h news window 24 times a day burns API calls without changing the score much.
I settled on 5 runs at 6am, 10am, 2pm, 6pm, and 10pm Central. That catches the US market open, midday, US close, evening news, and the late night Asian market window. Enough refreshes to feel current. Not so many that I’m paying for redundant work.
If you build your own version, start with twice a day and bump it up only if you feel like you’re missing things. The cost of running more often is linear. The information gain is not.
The prompt is most of the magic
The n8n workflow is mechanical. The Claude prompt is where the actual analysis happens. Here’s the shape of the per-coin prompt (simplified):
“` You are a crypto analyst writing for the holder personally. Analyze sentiment for {coin name} ({symbol}). Recent 24h news ({n} items): {numbered list of titles + summaries} Current: ${price} (24h {change}%). Holder owns {quantity}.
Return ONLY valid JSON: { “score”: <-100 to 100>, “verdict”: ““, “action”: ““, “recommendation”: “
Two things matter in that prompt:
- “Writing for the holder personally.” That single phrase changed the output a lot. Without it, Claude writes like a CoinDesk summary. With it, Claude writes like a friend who knows I own the thing.
- “Return ONLY valid JSON.” Required. Without this constraint, Claude wraps the JSON in markdown code fences and breaks the parser node. With it, the output drops cleanly into Airtable.
The “action” field is the one I treat with the most caution. It’s not a signal. It’s a starting point for a conversation with myself before I do anything. Same idea as the AI brainstorming panel, where the verdict is the prompt for my own thinking, not a substitute for it.
The mistakes I made building it
A few that are worth flagging if you build your own:
- First version dumped ALL the news into one giant prompt. Claude tried to score everything at once and the output was mush. Fix: per-coin buckets, one prompt per coin, then a separate macro pass.
- First version used a CryptoPanic API. It worked, then it rate-limited me, then it broke. Swapped to three free RSS feeds. More reliable, zero auth.
- First version had no dedupe. Same news items got scored on every run. I added a “Get Existing News” Airtable lookup plus a filter node so only NEW headlines hit the AI. This cut costs roughly in half.
- Forgot
executeOnceon the RSS reads. Items started multiplying down the chain. n8n quirk: any one-shot node like RSS or HTTP fetch needsexecuteOnce: trueor you get exponential item count.
The last one took me longer than I want to admit. If your n8n workflow ever produces way more items than makes sense, check executeOnce first.
Where the scores land
Three Airtable tables fill up per run: Coin Snapshots (one row per coin per run, with score/verdict/action/drivers), Macro Pulse (one row per run, portfolio-wide vibe), and News Items (deduped headline log). An Airtable Interface reads from those tables and Alpha the wolf, the mascot, is the one nodding at you when sentiment turns bullish.
Same pattern as my blog analytics workflow: n8n does the work, Airtable holds the state and powers the interface, AI handles the scoring. Reuse the pattern, change the inputs.
How to build your own
The short version, if you want to recreate this on your own n8n:
- Get your held coins into Airtable (or any data source)
- Pull live prices from CoinGecko’s free
/coins/marketsendpoint - Read crypto news from a couple RSS feeds (CoinTelegraph, Decrypt, BeInCrypto work well)
- Bucket news per coin in a Code node (substring match on coin name + symbol)
- POST to Anthropic’s
/v1/messageswith a structured prompt that demands JSON - Parse the JSON and write the row to Airtable
- Run the call once more across all coins for a macro vibe
- Schedule the whole thing 2-5x/day
It’s an afternoon build if your credentials are ready. If you want to go deeper on n8n + Claude, AI Automation Society Plus (affiliate link) is the Skool community I hang out in for this exact intersection. It’s where n8n and Claude people overlap, which is rare in general AI groups.
Frequently asked questions
Can AI predict crypto sentiment from news?
AI can read crypto news and score the vibe in seconds, but it cannot predict prices. What a crypto sentiment analysis tool does is summarize and quantify the tone of recent headlines, giving you a peripheral pulse instead of a trade signal. Use it as context, not as a buy/sell trigger.
How much does it cost to run a crypto sentiment workflow in n8n?
About $4 per month if you use Claude Haiku 4.5 and run it 5 times a day across 5 coins. Each run costs roughly $0.025 in API tokens. CoinGecko, RSS feeds, and Airtable’s free tier are all $0. If you’re already self-hosting n8n on a small VPS (~$8/month), this workflow rides for free on top.
Why use Claude Haiku instead of GPT-4 or Claude Sonnet?
Haiku 4.5 is the cheapest current Claude model and the sentiment scoring task doesn’t need a bigger brain. Claude Haiku handles structured JSON output reliably, follows the “return only JSON” constraint, and costs roughly 1/10th of Sonnet. For a workflow that runs 5 times a day across multiple coins, the cost difference adds up fast. Use Sonnet only if you find Haiku’s scores are noticeably off, which I haven’t.
What RSS feeds work best for crypto news?
CoinTelegraph, Decrypt, and BeInCrypto cover the bulk of major crypto news between them. All three are free, no auth, and the feeds are stable. CryptoPanic also has a free tier but the rate limits bit me, so I dropped it. Three feeds is enough for breadth without dragging in too much noise.
Is this safe to use for actual trading decisions?
No. This crypto sentiment analysis tool is built for peripheral awareness, not signal generation. It reads news, it doesn’t read order books, on-chain flows, or whale wallets. Use it to know what conversations are happening around your coins. Make real trading decisions with a much wider information set, and ideally with a real strategy.
Can I use this with other Airtable dashboards I already have?
Yes. The workflow is decoupled. It writes three flat tables and any downstream interface (Airtable, a custom dashboard, a cost-calculator-style page) can read from them. Swap in your own holdings table at the top and your own UI at the bottom.
The takeaway
A crypto sentiment analysis tool built in n8n with Claude Haiku gives you a daily news-driven pulse on your holdings for about $4 a month, with full control of the data and zero subscription to a paid signals service. It surfaces what’s being said about your coins so you don’t have to read every headline yourself. It does not predict prices, replace strategy, or give you an edge over the market. It buys back the hour every night you’d spend catching up.
I’ve got more n8n + Claude workflow blueprints on the way, including the JSON for this one. Subscribe if you haven’t yet and you’ll get them as they drop. Go forth and automate the boring research.
How would you change or improve this system?
