Prerequisites
Cryptohopper MCP configured in an MCP client. Claude desktop or Claude Code work best for conversational use, see setup overview.
Explorer tier or higher strongly recommended. Research conversations tend to pull candles, which require historical data — see subscription tiers.
A loose sense of what you'd want to ask. The assistant gets more useful as you learn which questions work.
Setup steps
Open your MCP client and confirm the Cryptohopper MCP is connected
Check the tool panel — the MCP should show as active before you start.
Run five calibration questions
Ask each, review the answer, note what worked and what didn't. After these five you know whether your client, model, and MCP setup are healthy.- What's the current BTC/USDT price on Binance, and how does it compare to Coinbase and Kraken?
- Run a quick TA on ETH/USDT on Binance, 4h timeframe.
- Scan the top 50 pairs on Binance by 24h volume and tell me the three most interesting movers.
- Compare the realised volatility of BTC and ETH over the last 30 days.
- If I wanted to buy $20k of SOL right now, which exchange would give me the best fill?Build a persistent context block
Drop this into the conversation at the start of each session, or use your client's persistent memory feature. It cuts 50% of clarifying questions and makes every answer tighter.Context about me as a crypto user:
- I trade primarily on Binance.
- My watchlist: BTC, ETH, SOL, AVAX, ARB, OP, LINK, AAVE, UNI.
- I mostly hold spot; I run grid bots on SOL and AVAX.
- I care about drawdown risk more than missing upside.
- I don't trade perpetuals.
When answering, default to the Binance exchange unless I specify otherwise. When running TA, default to 4h timeframe. Prefer concrete numbers over adjectives.Develop a set of go-to question patterns
Define a few reusable shortcuts — for example: "Give me the daily snapshot" (your watchlist digest), "Check {pair}" (TA + liquidity + volume check), "What's worth looking at today?" (top-movers + unusual-volume scan). See prompt patterns that work with the Cryptohopper MCP.
Keep asking follow-ups
The research-assistant mode is conversational — never accept the first answer as final. Push with: "How confident are you in that read?" / "What would change your mind?" / "Is there anything in the data you haven't mentioned?" The model gives better answers with pushback than without.
Save the conversations and prompts that worked
When a session produces a prompt or workflow that worked unusually well, name it and save it. Over months, the library is more valuable than any single session.
What a good session looks like
A realistic three-turn research conversation:
Speaker | Message |
You | What's worth looking at on Binance today? |
Assistant | Scanned the top 50 by volume. Three flags: ARB up 8.2% on 2.4× normal volume — real move, 4h trend confirmed up. SEI up 5.1% but volume flat — likely sympathy move, lower conviction. LINK sideways but volume 3× normal — no price move yet, worth watching for a break. |
You | Focus on LINK. Is there anything in the orderbook that explains the volume? |
Assistant | Pulled Binance LINK/USDT book. Depth healthy on both sides, no unusual walls within 2% of mid. The spike has already traded through — not visible in the current book. Worth checking on-chain activity or news for LINK. |
You | Is Binance the right venue for LINK today, or should I look elsewhere? |
Assistant | Checked LINK/USDT on Coinbase, Kraken, and OKX. Prices within 5bp of each other. Binance has the deepest book for a $50k+ order; Coinbase is close behind. For smaller orders, any of the four works. Spread is tightest on Binance. |
Cost profile
Research-assistant sessions are the most variable-cost workflows — it depends on what you ask.
Session shape | Approx call units |
Light ("what's the price?") | 5–15 |
Typical research session | 30–80 |
Heavy (multi-pair TA + multi-venue depth) | 100–300 |
Explorer's 30,000 calls/week supports a heavy session every day with quota to spare. On Pioneer, stick to ticker-only questions. See rate limits explained.
Troubleshooting
The assistant gives confident-sounding but vague answers
Require concrete numbers: "always include at least one price, percentage, or ratio per claim you make." Vague output is a prompt problem, not a model problem.
It re-derives the same setup every session
Add the context block from step 3. The assistant should know your watchlist, preferred exchange, and risk preferences without being told each time.
It pulls heavy data when light data would do
Be explicit when the question is cheap: "answer this from tickers alone." The model is generally good at picking the right tool but sometimes over-fetches on complex questions.
Answers differ significantly between sessions for the same question
The data changes run-to-run, which is correct. If the framing differs wildly, the prompt is under-specified. Encode the output shape: "always answer TA questions in this order: trend, momentum, key levels, one-line summary."
The assistant speculates about things the data can't show
When it starts talking about "the narrative" or "what the market thinks", pull it back: "only make claims that can be supported by the data you pulled; flag any interpretation as speculation."
You want the assistant to execute, not just research
See combine MCP + Cryptohopper Trading API for end-to-end agents (coming soon) and research via MCP, execute via REST API. The research assistant is the foundation; execution is a layer on top.
The client doesn't support the conversational style well
Some clients (pure IDE agents) prefer short instructions, not back-and-forth. For conversational research, Claude desktop and Claude Code are usually the best experience. See setup overview.