ai

Building a TradFi analysis skill with Binance Agent OS - by a guy who had never written a trading bot.
THE PROBLEM I ACTUALLY HAVE
I am a 9-5 Business Analyst at a securities firm in Hanoi. I understand markets from the business side, not the trading side.
Here is my problem. Vietnam Securities market closes at 14:45. The US market opens at 21:30 my time, and the interesting moves happen around 03:00, when I am asleep. Every morning I open four tabs and manually reconstruct what happened.
I wanted an agent to do that for me. Not one that trades - one that tells me what moved and by how much, in numbers I can check.
WHAT I BUILT
A skill called tradfi-trading-skill. One job: every morning, pull the overnight tape for the instruments I care about, compute real volatility, and tell me what was unusual.
The thing I did not expect is that Binance carries TradFi perpetuals - $NVDA , $TSLA and others - trading 24/7. So I can watch US equity moves through the same connection as crypto, at 3am, without a US broker.
SETUP: 15 MINUTES, NO CODE
Step 1 - Connect Agent OS.
In Claude, add the Binance MCP server and authenticate. Then the part that matters: set permissions to market-data read-only. My skill never places an order, so I did not grant trade access. If it cannot trade, it cannot mis-trade.
Step 2 - Test with something trivial.
What is the current XAUUSDT price and 24h change on Binance?
It returned 4,626.26, up 0.23%. Connection confirmed. Do this before you build anything.
Step 3 - Ask for the real thing
Using the Binance MCP, pull the last 10 daily candles for the NVDA TradFi perpetual. Compute the average daily range, then tell me which day was unusual and why. Show me the raw numbers.
Step 4 - Ask the AI ​​to package the analysis into a skill/skillset
THE RESULT
After I called tradfi-trading-skill on NVDA, claude response with this:
● NVDA perp last 220.74; average daily range about 2.94% of price
● One candle stood out - open 213.75, high 220.45, low 203.80, close 219.46
● That is a 7.79% range, over 2.5x a normal day
● Volume 2,861,367 against a 425,031 baseline - 6.7x
● And it closed having recovered 94% of the day range
That was the earnings reaction. A sharp drawdown almost entirely bought back within the session. I got the whole shape of it in one answer, from data I can re-pull and verify myself.
Then I asked it to compare against BTC and got something I did not expect:
BTC is only about 1.1x more volatile per day than an NVDA perp right now. As someone who designs risk warnings for a living, that reframed something. But look at the trade count - 5 million against 37 thousand. Similar volatility, completely different liquidity. Those two things need different interfaces, and I had never seen it stated in numbers before.
FOUR MISTAKES THAT I MADE WHILE BUILDING ON AGENT OS
1. I assumed the MCP had news. It does not.
There is a tool called analysis.getTokenAiReport. It returns success: true and null. I built a whole plan around it before testing it. Test your tools before you design around them.
2. I trusted an article over the order book.
An article dated that same day said BTC was trading near $60,000. The Binance ticker said $78,444. The article was $18,000 wrong. Auto-generated market content is everywhere now. The exchange is the source of truth.
3. I thought the MCP could post to Binance Square. It cannot.
Publishing is a separate skill with a separate key. That is good design - the posting key cannot touch funds - but I lost an hour assuming otherwise.
4. I called a tool before it was loaded.
Tools load on demand. If you get a "not loaded" error, that is not a permissions problem. Ask Claude to find the tool first.
WHAT IT CANNOT DO
● It will not trade for you. Every consequential action needs your confirmation, by design.
● It has no news feed. Pair it with web search.
● The most recent candle is always incomplete, so its volume looks artificially low.
● It reports what happened. It does not know what happens next, and neither do us.
FINAL
I did not build a trading bot. I built something that removes twenty minutes of manual tab-checking each morning and replaces it with numbers I can verify.
What changed my mind was not the automation. It was that checking became cheap. When verifying a claim costs one sentence instead of ten minutes, you start verifying claims you would previously have just believed. For someone who designs financial interfaces, that is the more useful habit.
I am a business analyst, not a developer. If I could build this in an afternoon, so can you.
@CZ @Yi He @Binance Square Official @Binance Blog @Binance Academy
Not financial advice. All figures pulled live from the Binance API and reproducible with the prompts above.
#AI #AgentOS