Build with agents
Agents can already read a user’s exchange account and place orders. What they lack is evidence: where today’s readings stand against each asset’s own history, and the historical outcomes of similar readings. CRYPTYX supplies that layer through a remote MCP server and a REST API. This guide shows where it fits, the flows to run, and how to read what comes back. No trading background is assumed.
Where CRYPTYX sits
Asks questions, sets the rules, approves any order.
Claude, ChatGPT, Grok, or your own framework. Calls tools, composes the answer.
Coinbase, Kraken, Binance, Hyperliquid tools. Reads holdings, executes approved orders.
Evidence layer. Receives asset symbols only and returns readings and their history.
The agent reads holdings from the exchange, sends only the symbols to CRYPTYX, and combines both. CRYPTYX never receives balances, position sizes or account identifiers, and never places orders. Responses describe what the data shows and its historical outcomes. They do not tell anyone to buy or sell; the user’s rules or your bot’s logic decide that.
The engine, layer by layer
Each layer is built from the one below. Most users start at the top with a plain read of one asset; developers and quants can go down to individual readings and build their own conditions.
| Layer | What you get | Surface |
|---|---|---|
| THE STUDY | One asset, two horizons: readings that stand out and the historical outcomes of similar readings | get_coin_study · /api/v1/study |
| Episode study | Every prior instance of a move and its forward returns | get_episode_study · /api/v1/episode-study |
| Readings | Test one reading: every past episode and the full return profile at 8 horizons | analyze_metric · /api/metrics/slicer |
| Combinations | Combine 2–4 readings across factor families and test them together | analyze_metrics_composite · /api/metrics/slicer/composite |
Connect
MCP (any agent client)
Add the remote server and sign in through your client’s OAuth flow. Free accounts get THE STUDY, the episode study, coverage and data freshness with daily limits; scans and backtests need a Pro account.
https://mcp.cryptyx.ai/api/mcp
- Claude Code:
/plugin marketplace add cryptyx-ai/agents, installcryptyx, then/mcpto sign in. The plugin adds skills that route questions to the right tools. - Claude, ChatGPT, Grok and others: add the URL above as a custom remote MCP connector and sign in.
- Skills source: github.com/cryptyx-ai/agents (MIT). Readable as agent instructions for any framework.
| Free tool | Calls per day |
|---|---|
| get_coin_study | 10 |
| get_episode_study | 3 |
| check_coverage | 30 |
| get_data_liveness | 60 |
| get_cryptyx_overview | unlimited |
REST (bots and backends)
The same reads are plain HTTPS endpoints. Authenticate with an API key, or pay per call with x402 (USDC on Base) and no account. Full reference in the API reference and x402 guide.
# THE STUDY for one asset curl -H "Authorization: Bearer $CRYPTYX_API_KEY" \ "https://cryptyx.ai/api/v1/study?asset=SOL" # Scan up to 25 assets, with what changed over 7 days curl -H "Authorization: Bearer $CRYPTYX_API_KEY" \ "https://cryptyx.ai/api/v1/heat-check?assets=ETH,SOL,DOGE&compare_days=7" # What stands out on up to 10 assets curl -H "Authorization: Bearer $CRYPTYX_API_KEY" \ "https://cryptyx.ai/api/v1/notable?assets=BTC,ETH,SOL" # Is an asset covered? When did each data source last update? (free) curl "https://cryptyx.ai/api/v1/coverage?assets=SOL,BONK" curl "https://cryptyx.ai/api/v1/liveness"
Flows to run
Each flow is a prompt a real user would type, the tools a well-behaved agent should call, and what a good answer contains. Run them in each client you care about. They also make a useful regression suite for your own agent.
1. Research one asset
Anyone, no exchange needed“Report where SOL stands right now against its own history, and the historical outcomes of prior instances like it.”
Expected tools get_coin_study(asset: SOL)
- ✓Separates the tactical (7–14 day) and structural (60–180 day) pictures.
- ✓Names the readings that stand out in plain words, with their z-scores.
- ✓Quotes the historical outcomes with the number of prior instances and the base rate, not as a prediction.
2. Check my holdings
Exchange connected in the same agent“Pull my Coinbase holdings. Which of my holdings stand out, and why?”
Expected tools exchange balance tool → get_coin_study on each holding
- ✓Only symbols reach CRYPTYX; balances stay with the exchange tool.
- ✓Uncovered assets are named as not covered, not skipped silently.
- ✓Explains why each asset stands out using the returned readings, not general knowledge.
3. What changed
Recurring check“Run the study on ETH, SOL and HYPE. What changed since last week?”
Expected tools get_coin_study on each asset, re-run by your own agent
- ✓Opens with readings whose tag changed since last week.
- ✓Says plainly when nothing changed.
4. Test your own idea
Developers and quants (Pro)“On BTC, test strong 60-day trend efficiency together with negative funding. Show every horizon against the base rate.”
Expected tools get_featured_metrics → analyze_metric → analyze_metrics_composite
- ✓Reports episodes, trigger rate, median and mean, hit rate, Sharpe and drawdowns at each horizon.
- ✓Leads with the difference from the base rate (median_vs_base, hit_rate_vs_base).
- ✓Relays the response’s warnings verbatim, including small samples.
- ✓Shows all horizons rather than only the best one, and says how many variants were tried.
5. Feed a bot
Your own code“Every morning, read THE STUDY for my watchlist and log any reading that crossed into its top or bottom 5%.”
Expected tools GET /api/v1/heat-check (daily) → GET /api/v1/study per asset → your rules
- ✓Checks /api/v1/liveness first and skips a run when data is stale.
- ✓Treats CRYPTYX output as an input to its own rules, with its own risk limits.
- ✓Keeps the decision and any order approval outside CRYPTYX.
Reading the evidence
Every historical number comes with its sample. A few rules keep an agent, or a bot, from over-reading it:
- Compare with the base rate before calling anything an effect. The response carries both.
- Prefer medians: a few large moves pull the mean.
- Under about 15 episodes, treat the result as anecdotal.
- Every threshold or horizon tried is another draw. A real effect holds on nearby thresholds and similar assets.
- Fields named
warningsare written for relay to the user; pass them on as they are. - History is history. Nothing returned is a forecast or advice.
Glossary for non-traders
- Reading
- One measured quantity for one asset on one day, such as funding paid by leveraged traders or realised volatility over 30 days. CRYPTYX tracks 470+ of them across about 180 assets.
- z-score
- How far today's reading sits from the same asset's own past-year norm, in standard deviations. +2 means far above the asset's normal; −2 far below.
- Percentile
- Where today sits in the asset’s own history (own-history percentile) or against all other assets today (market percentile).
- Horizon
- The window a reading looks over. Tactical readings cover 7–14 days, structural readings 60–180 days. They answer different questions and are never blended.
- Factor family
- A group of related readings: trend, trend efficiency, volatility, flows, futures positioning, order book, options and correlation.
- Episode
- A past occasion when a condition held, counted once even if it held for several days in a row.
- Forward return
- What the price did over the next N days after an episode (7, 30, 90 and more).
- Base rate
- What the price did over the same N days from every day in the same period, whatever the readings. The comparison point for any condition.
- Hit rate
- Share of episodes followed by a positive return. Only meaningful next to the base rate: an asset that rose on 58% of all days makes a 60% hit rate unremarkable.
- Sharpe
- Return per unit of volatility, annualised. A rough quality measure for a stream of returns.
- Drawdown
- The worst fall from entry along the way, even if the period ended up.
- Setup
- Two to four conditions across different factor families holding at once. CRYPTYX publishes a catalogue of setups with evidence tiers.
- Signal
- A condition the engine evaluates daily with a tested track record.
Reporting what you find
When a flow misbehaves, the useful report is: client and model, the prompt, the tools actually called (in order), the part of the answer that was wrong or missing, and the response field it should have used. Send it to support@cryptyx.ai.
CRYPTYX provides historical market statistics. Nothing here is an offer, a solicitation or investment advice, and historical results do not predict future results. You are responsible for your agent’s actions and for any orders you approve. Digital assets carry substantial risk of loss.