Heracles walked into the valley of Nemea to kill a lion nobody could wound. He won by not fighting it head on. Nemea does the same for a portfolio: crypto's "invincible" drops are survivable when something is watching for you and tells you calmly what is going on.
Every crypto tool is a screen you have to be watching. Nemea is the one that writes to you when you are not.
Replayed against a real year of CoinMarketCap daily prices, it sent 94 alerts and held back 396 more with its own cooldowns. On the single crash day in that year it alerted on the one coin that crossed the threshold and stayed quiet on the three that fell 7%, 12% and 14.6% without crossing it. Numbers and method: docs/evidence/replay.json, reproducible with pnpm replay.
That restraint is the product. Before an alert reaches you it asks one question a bare price alert cannot: is this just your coin, or is CoinMarketCap's own category data showing the whole sector down. "Your portfolio dropped 9%, and 64% of that is Smart Contracts falling 8% on average" is a different alert than "ETH is down". Nemea also watches for a stablecoin losing its peg, a volume anomaly and a stale quote, and offers optional protective steps through your own wallet. Every alert has an Explain like I'm 5 button that turns a panic moment into a learning moment.
Not a trading bot. Insurance against being offline. Non-custodial, no keys, no trading. It is not a signal service and not built for people hunting alpha. It is for people who bought crypto and want to keep it.
- CoinMarketCap rate limits are real. Nemea reads your key's plan and adapts: for a small portfolio on the free plan the stablecoin peg check runs every 4 minutes, not every minute.
/statusshows the real cadence and every real API call. - Not financial advice. Every alert carries the disclaimer. Every protective action is a link you approve in your own wallet.
- Non-custodial by design. Nemea reads wallets by public address. It never asks for, stores or handles keys or seed phrases, and
pnpm check:claimsfails the build if source code starts to.
Details, numbers and the command that checks each claim: docs/HONEST_LIMITS.md.
Live demo: https://nemea.unitynodes.com. Demo account: none needed. Press Start as guest, then Load a sample portfolio, then Try an alert.
Prerequisites: Node 22+, pnpm 10.
pnpm install
pnpm dev:stack # API on :4000 on a fake CoinMarketCap (development only, NOT real market data)
pnpm dev:web # web on :3000Open http://localhost:3000, press Start as guest, Load a sample portfolio, then Try an alert and Explain like I'm 5. If port 3000 is taken, run the web app on another port and start the stack with WEB_ORIGIN=http://localhost:<port> pnpm dev:stack; the API refuses requests from any other origin.
Measured on 2026-09-21 from a fresh git clone with a warm pnpm cache: pnpm install --frozen-lockfile 3 s, web app ready 11 s, and the whole journey (guest, sample portfolio, simulated alert, Explain, swap link) 1.6 s over HTTP. A cold install takes longer; time it on your own machine before trusting "under 5 minutes".
With real keys, copy .env.example to .env, fill it in, and run pnpm dev:api, pnpm dev:web, pnpm dev:bot. Then check that your keys work with the gates:
pnpm gate:a # CoinMarketCap: plan, quotes, global, categories, history, budget
pnpm gate:b # wallet read on Ethereum, Base, Arbitrum and CMC matching
pnpm gate:c # Telegram: send a test message
pnpm replay # replay the real alert engine against a real year of CoinMarketCap historyA gate that could not run exits 2. A gate that examined nothing never exits 0.
| Alert | Fires when | Default | Data |
|---|---|---|---|
| Price drop | a held coin falls more than X% in 1 h or 24 h | 8% / 15% | quotes percent_change_1h, percent_change_24h |
| Below cost basis | price crossed below what you paid within the last day | if you entered a cost basis | quotes |
| Near low | price within X% of the all-time low (or the 365-day low when the plan has no all-time low, labelled as such) | 10% | price-performance-stats or historical |
| Stablecoin depeg | a held USDT, USDC, DAI or USDe trades below the floor | $0.98 | quotes |
| Stablecoin volume | volume at least 10x the previous day | 10x | quotes volume_change_24h |
| Volume spike | volume at least 5x the previous day | 5x | quotes |
| Volume dry-up | volume down 70% while the price falls (a proxy for thin liquidity) | 70% | quotes |
| Category rotation | a category you hold falls X% and at least 5 points worse than the market | 10% | categories, global metrics |
| Portfolio drop | portfolio down X% in 24 h, with a breakdown by category ("because Layer 1 is down 18%") | 10% | quotes, info tags, categories |
| Tokenised asset drift | a tokenised stock or commodity you hold trades away from what every other version of that asset trades at | 3% | real-world-assets map and quotes |
Alert fatigue is a design constraint: at most 3 non-critical alerts a week (you can lower it), cooldowns per alert, a roll-up so one crash is one alert, and no alert at all on stale or missing data.
Delivery: web dashboard, Telegram, email (critical immediately, everything else in one daily digest), browser push. Protection levels: 1 alert only (default), 2 alert plus swap links for Uniswap and 1inch that you approve in your own wallet. Level 3 (auto-swap) is not built.
| Endpoint | Used for | When |
|---|---|---|
/v3/cryptocurrency/quotes/latest |
prices, 1 h / 24 h / 7 d / 30 d change, volume and volume change | every minute on a large plan (every 4 minutes on the free plan), batched to 100 coins per call |
/v3/cryptocurrency/quotes/historical |
"similar drops recovered in X days", peg history, 365-day low | on demand (Explain), cached 6 h |
/v1/global-metrics/quotes/latest |
market-wide move, dominance | every 15 min on a large plan |
/v2/cryptocurrency/info |
tags with groups, contract addresses per chain (wallet tokens are looked up by symbol, then matched by exact chain and contract address) | once per coin, then stored; once per wallet import |
/v1/cryptocurrency/categories |
category averages for "because Layer 1 is down X%" and rotation | every 30 min on a large plan |
/v2/cryptocurrency/price-performance-stats/latest |
all-time low, when the plan allows | in the background for your largest holdings, cached 24 h; falls back to historical when the plan does not allow it |
/v1/key/info |
plan limits, so the polling budget is computed, not assumed | at start and hourly |
/v5/real-world-assets/map |
which tokenised assets exist and which actually have wrappers | with the RWA refresh, 0 credits |
/v5/real-world-assets/quotes/latest |
the average price across every tokenised version of a real asset, and each wrapper's own price and CoinMarketCap id | every 4th category run; 282 assets and 829 wrappers cost 5 credits |
The /status page lists the last real calls with time, endpoint, HTTP status, credits and latency.
Every number in Nemea comes from calls like this one, sent by packages/cmc-client/src/client.ts. Captured by pnpm gate:a on 2026-09-21 with a Basic key (the key is redacted in the file). Bitcoin, Ethereum and USDC in one request cost 1 credit.
curl -s "https://pro-api.coinmarketcap.com/v3/cryptocurrency/quotes/latest?id=1,1027,3408&convert=USD" \
-H "X-CMC_PRO_API_KEY: $CMC_API_KEY"const client = new CmcClient({ apiKey: process.env.CMC_API_KEY });
const { quotes, missing } = await client.getQuotes([1, 1027, 3408]);The response, trimmed to Bitcoin and to the fields Nemea reads (the full body with all three coins is in docs/evidence/gate-a-sample-call.json, the receipts of all 10 calls of the run in docs/evidence/gate-a-cmc.json):
{
"status": {
"timestamp": "2026-09-21T17:36:24.903Z",
"error_code": "0",
"error_message": "",
"elapsed": 4,
"credit_count": 1
},
"data": [
{
"id": 1,
"name": "Bitcoin",
"symbol": "BTC",
"cmc_rank": 1,
"last_updated": "2026-09-21T17:35:00.000Z",
"quote": [
{
"symbol": "USD",
"price": 85884.0831679611,
"volume_24h": 51534296437.904594,
"volume_change_24h": 142.2458,
"percent_change_1h": 0.17164259,
"percent_change_24h": 5.68945324,
"percent_change_7d": 9.03108265,
"market_cap": 1725200471304.2444,
"last_updated": "2026-09-21T17:35:00.000Z"
}
]
}
]
}Captured on 2026-09-22 from the live site on real CoinMarketCap data with CONFIRM_REAL_DATA=1 BASE_URL=https://nemea.unitynodes.com node apps/web/scripts/screenshots.mjs. The script refuses to run against the development fake, so these cannot be mock data. Every width is in docs/screenshots/.
A portfolio alert that says where the drop came from. Not "ETH is down" but which sector carried it, from CoinMarketCap's own category averages.
Explain like I'm 5. Deterministic, no language model: the same alert in plain words, including what Nemea could not check.
Protective action, level 2. A swap link you open and approve yourself. Nemea never signs anything.
The status page. The plan, the schedule, the track record, and the log of every real CoinMarketCap call.
Every demo alert in this README and on the live site can be a simulation, labelled as one. To show the rule engine actually holds up, pnpm replay runs the unmodified production engine (packages/alerts) against a real year of CoinMarketCap daily closes for BTC, ETH, SOL, DOGE and USDC (2025-09-24 to 2026-09-22), one coin at a time, with the same caps and cooldowns a live portfolio gets. Full method, every date fixed to a UTC calendar day for reproducibility, and all 94 alerts: docs/evidence/replay.json. The same numbers are on the live site: nemea.unitynodes.com/status, under "Tested against a real year of history".
The worst day in the window was 2025-10-11. BTC fell 7.2%, ETH 12.2%, SOL 14.6%, DOGE 22.3%. Only DOGE crossed the default 15% threshold, and only DOGE got an alert. That is the point of thresholds: three real double-digit drops that day correctly stayed quiet.
Over the year, the engine would have sent 94 alerts (1 price drop, 93 near-a-low notices) and held back 396 more with cooldowns alone, no weekly cap needed. Naive daily checking without cooldowns would have been roughly 5 times noisier.
This is a measurement of the shipped engine on real prices, not a backtest of a trading strategy: recovery after an alert is not the point and is not claimed anywhere. depeg and price_drop_1h cannot be exercised this way (no hourly bars that far back on this plan, no volume in quotes/historical); portfolio_drop and category_rotation need a real multi-coin portfolio and category data, not a single-coin replay.
CoinMarketCap tracks 7,942 real-world assets: tokenised stocks, gold, treasuries. Each one has several wrappers from different issuers, and they are supposed to track the same thing. Usually they do. Sometimes one does not, and if you hold that one, nothing tells you.
Nemea watches the gap. Your wrapper's price against the average across every tokenised version of the same asset, from /v5/real-world-assets/quotes/latest. That is the depeg rule pointed at a different anchor: a stablecoin should be $1, a tokenised NVDA wrapper should be what tokenised NVDA costs everywhere else. Alert, cooldown, weekly cap and Explain like I'm 5 are the same ones the rest of Nemea uses.
On a live run on 2026-09-22 it read 282 assets and 829 wrappers for 5 credits in 1.4 seconds, and 16 wrappers were sitting 3% to 25% away from their asset's average, the widest being a Hyperliquid TSLA wrapper at 20.8% below.
The trap, and what it cost. Seventeen more wrappers looked like enormous dislocations and are not: GRAMS prices silver per gram while the asset average is per troy ounce, so it reads 93% low, and XAGX reads 112% high for the same reason. Nothing in the response says which unit a wrapper uses. A naive threshold would alert on those every single poll forever, which is exactly the noise Nemea exists to avoid, so anything past 25% is treated as a different unit rather than a market move and is recorded, not sent. packages/alerts/test/rwa.test.ts holds that line.
Getting here cost an afternoon of wrong turns that are worth passing on: the RWA family is under /v5/real-world-assets/, guessable from no sibling, and an unrouted CoinMarketCap path answers 200 with "The system is busy, please try again later!" rather than 404, so a wrong guess looks exactly like an outage. Then quotes/latest rejects a whole batch of ids if even one asset has no wrappers, and there is no skip_invalid to soften it, so the free map call has to filter on has_tokens first. All of it is in docs/API_FEEDBACK.md.
- ...a price alert? A price alert is one number crossing one line for one coin. Nemea builds each alert from your holdings: it knows a stablecoin should be worth $1, that your portfolio fell mostly because one category fell, and that a quote is stale. It also tells you when not to worry, using the coin's own history.
- ...a portfolio tracker? A tracker shows you what you have. Nemea is silent until something matters and then explains it.
- ...a trading bot or signal service? They add risk and demand attention. Nemea never trades, never predicts, and never touches your funds.
- ...checking the app more often? Nemea is for the moments you are offline.
- Watching a whole portfolio for a few credits.
quotes/latesttakes 100 coins per call and returns price, 1 h / 24 h / 7 d / 30 d change and volume change together, so a portfolio check is one credit, not one per coin. - Peg and volume rules without a market feed of our own. Stablecoin price and
volume_change_24hfrom the same call are enough for the depeg and volume-anomaly alerts. - Context for a move. The categories endpoint has a live 24 h average per category (360 of 360 had one on 2026-09-21; Ethereum mapped to Smart Contracts at +4.4%), which is what lets an alert say "the whole category moved" instead of a bare number.
- Wallet import that matches the right token.
inforeturns a contract address per chain, so a token found in a wallet is matched by exact chain and address, not by a ticker that anyone can copy. - A polling budget that is computed.
/v1/key/inforeturns the plan limits, so the cadence is planned from the real credits and the/statuspage shows it.
Real friction from building this, so the next builder does not lose the time:
- Endpoint versions. v1 and v2
quotes/latestandquotes/historicalare deprecated; v3 returnsdataandquoteas arrays, older versions return keyed objects. The parser accepts both. - The free plan cannot support a one-minute peg check. 15,000 credits a month against roughly 43,200 minute-polls. Cadence is planned from
/v1/key/infoand shown honestly. - No all-time low in quotes. It exists only in
price-performance-stats, which the pricing matrix does not list for Basic or Builder. Nemea falls back to a labelled 365-day low. - Info by contract address can return another chain's
platform.token_address; wallet tokens are matched oncontract_address[]by chain and address instead. - No liquidity or order-book data in the basic data, so "sudden liquidity drop" is a volume dry-up proxy and says so.
infoby contract address is fragile. Any unknown address in a batch makes the whole call HTTP 500 even withskip_invalid=true, and checksum-cased addresses return 400 or 500. Wallet tokens are matched throughinfo?symbol=(unknown symbols return an empty list) and then by exact chain and contract address.- Invalid ids in
quotes/latestare silently omitted, but a request where every id is invalid is HTTP 400 withcredit_count0.status.error_codeis a string on some endpoints and a number on others. - Tags are not ordered by relevance. Bitcoin's first tag that matches a category is an investor portfolio bucket ("Coinbase Ventures Portfolio"), Chainlink's is "Cosmos Ecosystem". About 65 of the 359 categories are investor, regulatory or estate buckets and most
*-ecosystemcategories are platforms, so Nemea usestag-groups, drops those, and explains a move through a thematic category (captured fixture and test:packages/alerts/test/categories.test.ts). - Category timestamps.
last_updatedon categories is a metadata date (348 of 359 are older than 30 days) while the averages are live, so it cannot be used as a freshness check. Nemea does not use it.
- Native mobile app
- More chains
- DEX-native protection (protective swaps inside the DEX instead of a link)
- Level 3 auto-swap through a pre-approved agent wallet, opt-in
- Smart-contract wallet sign-in
- Referral links once partner IDs exist
apps/api Express + Drizzle service, poller, delivery
apps/bot Telegram bot (grammY)
apps/web Next.js dashboard
packages/ shared-types, cmc-client, alerts, chain-reader
docs/ SCOPE, ARCHITECTURE, HONEST_LIMITS, DEPLOY, SUBMISSION
scripts/ gates, dev stack, claims check
More: architecture, scope and what was cut, deploy.
Status on 2026-09-22. "Not run" means exactly that.
| Claim | How it is checked | Status |
|---|---|---|
| Alert engine, cadence maths, CoinMarketCap client, API, bot, wallet reader, web helpers | pnpm test |
550 tests pass |
| Types | pnpm typecheck |
passes in every package |
| No key handling in source | pnpm check:claims |
passes (137 source files) |
| Fresh clone to first alert | clone, install, start, walk (see above) | verified |
| UI at 1280 px and 390 px | Playwright walk of the whole journey against the dev stack | verified on development data |
| Blockscout wallet read on Ethereum, Base, Arbitrum | live run on a public wallet | verified |
| Etherscan V2 free tier covers Ethereum and Arbitrum, not Base | live probe | verified |
| CoinMarketCap response shapes and error codes | CoinMarketCap docs and keyless live responses, captured fixtures in tests | verified against docs and keyless responses |
| Real CoinMarketCap calls with an API key | pnpm gate:a |
passed 2026-09-21 on a Basic key (15,000 credits/month, 50 requests/minute): 10 real calls, 11 credits, receipts in docs/evidence/gate-a-cmc.json, raw request and response in docs/evidence/gate-a-sample-call.json |
| Telegram delivery | pnpm gate:c |
passed 2026-09-21 with a real bot: token accepted, test message accepted by Telegram (message id returned) and confirmed on the owner's phone, and an over-limit message is rejected as expected. No evidence file is committed because it would contain a chat id. The link-code flow between the web app and the bot is covered by tests, not yet exercised end to end against the real bot |
| Tokenised-asset drift, end to end on the deployed site | added a real wrapper to a live portfolio and waited for the poller | verified 2026-09-22: a Hyperliquid TSLA wrapper at $300.00 against a $379.00 average across all tokenised TSLA produced one alert, with Explain like I'm 5, about three minutes after the holding was added |
| The shipped alert engine, replayed on a real year of CoinMarketCap history | pnpm replay |
run 2026-09-22: 94 alerts over 2025-09-24 to 2026-09-22 on BTC/ETH/SOL/DOGE/USDC, 396 more correctly held back by cooldown, evidence in docs/evidence/replay.json |
| Wallet import matched against real CoinMarketCap | pnpm gate:b |
passed 2026-09-22 on a public wallet: 79 tokens matched by chain and contract address across Ethereum, Base and Arbitrum, native ETH on all three, unmatched tokens skipped with a reason, evidence in docs/evidence/gate-b-wallet.json. The count moves between runs because the sample wallet trades. The gate reads public Blockscout, which sometimes times out on a wallet with thousands of balances; one failed run in three is normal and the gate reports it rather than passing quietly |
| Email and browser push | tests with stubs | not verified against Resend or a real push service |
| Deployed on the Unity Nodes server behind Cloudflare | docs/DEPLOY.md; curl https://nemea.unitynodes.com/api/health; each service survives kill -9 |
live since 2026-09-21; the live journey (guest, sample portfolio, simulated alert, Explain, swap links) was walked over https |
The gates write docs/evidence/*.json on a clean run. Commit those files and this table can change from "not run" to a link.
MIT. Built by Unity Nodes.