Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
Market anomaly feed
A scanner that scores real market series for statistically unusual moves and records each one as an event. Events are read through a paid x402 endpoint and through the market_anomaly alert rule kind.
What is scanned
| Processor | Series | Source | Cadence |
|---|---|---|---|
price |
Top 100 coins, 7 day hourly sparkline | CoinGecko markets table | hourly points |
tvl |
Solana DeFi TVL, last 120 days | DeFiLlama | daily |
stablecoin_supply |
Solana stablecoin circulating supply, last 120 days | DeFiLlama | daily |
stablecoin_peg |
Top 100 stablecoins vs $1 (USD pegs with at least $10M circulating) | buildStablecoins() |
spot |
funding |
Top 20 Hyperliquid perps, 7 days of funding history (8h-equivalent percent) | Hyperliquid fundingHistory |
hourly |
open_interest |
Top 20 Hyperliquid perps open interest | Hyperliquid, snapshotted each tick into market_anomaly_samples |
15 min |
Gas and whale processors are not part of this feed: gas is EVM-only and whale detection already lives in the wallet and Oracle surfaces.
Detectors
api/_lib/market-anomaly.js is pure (no network, no clock). For each series it builds overlapping lookback returns and scores the latest one:
- Modified z-score: median and MAD, scaled by 0.6745, threshold 3.5. Falls back to mean absolute deviation when MAD is 0, and refuses a constant history.
- EWMA deviation: lambda 0.2, threshold 3.
- Rate of change: the move in percent, gated by a per-series minimum.
Severity is the lower of the score rung and the move rung, so a large z-score on a tiny move is never reported. A series with fewer than MIN_POINTS (24) plus the lookback is skipped as insufficient_history; a series with no variance is skipped as flat_history. Every skip is logged with its reason and returned by the cron.
Events dedupe on event_key (processor:subject:metric:bucket), so a condition that persists writes one event per bucket, not one per tick.
Reading events
GET /api/x402/market-anomalies ($0.001 USDC, Solana or Base). Parameters: processor, severity (minimum, default medium), subject, hours (1-168, default 24), limit (1-200, default 50). Bad params are a 422 and a store outage is a 503; neither charges the buyer.
Alerts
Create a rule of kind market_anomaly (REST POST /api/alerts/rules, or create_alert_rule in packages/alerts-mcp). Optional filters: processors, min_severity (default high), subject. It takes no target mint. The alert runner delivers each new event once per rule.
Operations
Cron /api/cron/market-anomalies runs every 15 minutes (see vercel.json). It runs the six processors in parallel with per-processor error isolation and returns {scored, detected, written, skipped, skipped_detail, errors}; it is ok unless every processor failed. Open interest needs 29 snapshots (about 7 hours) before it scores. Tests: tests/market-anomaly.test.js, using series recorded from the live upstreams in tests/fixtures/market-anomaly/.