Kairos Signal: verifiable DePIN and physical-world data, with cryptographic receipts. Choose access based on the datasets, request volume, monitoring, delivery, and support your workflow requires. Check network and metric coverage before paying. Custom work and onward distribution require a separate scope or agreement where applicable. Provenance-first: every value carries its source, as_of, and a verify_yourself URL you can check against upstream. 447 symbols carry live data (318 networks + 129 Bittensor subnets), 288 of the networks first-party, read straight from the network's own API or chain, across 7,298 first-party series — every one returning the upstream endpoint you can call yourself. 30 more carry on-chain and market values only. Each value states its own age; stale is labelled stale. Verify free before you pay a cent. Browse every network we cover with its live metrics and the URLs to check them against.
The Partner Brief is generated daily (08:00 UTC) and published publicly at /briefs/ and GET /v1/brief for everyone, free; it is not a per-partner delivery. It covers capacity-vs-demand divergence, stale-dashboard catches, on-chain supply verifiability, exchange-flow extremes and holder concentration, each line sourced and independently checkable. See today's edition →
See live utilisation data for Akash, io.net, Storj, and ThreeFold — or query the cross-network comparison API (Design Partner and above; one canonical question per concept, and every cell carries its native metric, source and as_of; the concept and cell counts are in the response).
Paying in USDC instead? USDC top-ups need a current quote; two modes: unique amount (micro_decimal, no wallet signature, any sender) or verified wallet (sign a challenge once; only transfers from that wallet are credited).
What it computes, from the DAG manifold encoder's vectors (an untrained random-feature map of each tick, not correlations): volatility estimates (1h/4h/24h) from LightGBM models fitted on manifold features, Student-t VaR built on those estimates (its tail shape comes from a fitted model at 1h only; 4h and 24h use a fixed default and say model_available: false), a global regime label from two encoder-batch statistics (the batch half-norm ratio and the batch log top singular value, still exposed under the route names curvature and Lyapunov; the log top singular value grows with how many ticks a batch held, neither is a forecast, a risk measure or a trading signal, and neither is a Wilson-loop curvature or a Lyapunov exponent), a fragility index per symbol, and each symbol's nearest manifold neighbour with the pair's 24-hour price-divergence z-score. Those manifold vectors come from the research encoder, whose weights are an untrained initialization (a fixed random-feature projection of each input, with fresh layer-51 noise on each pass), and every answer says so in model_state. It is research output, not a forecast product and not a trading signal: the volatility model failed its skill test (below), and the pairs label (TRADE/HOLD/EXIT/WATCH/NEUTRAL) is a fixed threshold rule on manifold distance and price divergence that the skill test does not cover. No directional predictions.
Free beta. We will not charge for this until the model demonstrably beats a free EWMA/GARCH baseline out-of-sample. Skill test re-run 2026-08-10 on rebuilt price data (1,605 out-of-sample hourly forecasts, 39 symbols, Diebold–Mariano on QLIKE; model 2.20 vs EWMA 1.75, DM +5.42): the trained model did not beat the EWMA baseline, so it stays free. An earlier run the same day was voided when a price-scale defect was detected in the underlying candle table; it has been rebuilt from the exchange API and verified against an independent tick feed. Every /v1/shape answer carries the full verdict under calibration.skill_vs_baseline, and every answer from the four endpoints carries a summary of it beside model_state; this line will be updated if a future model passes.
Prepaid agent-credits rail for autonomous AI buyers. Credit-based ($1 = 1 credit). No card, self-service — $5 free credits on self-registration, enough to evaluate (limit 3 per originating IP per 30 days; past that a key is created at $0 and the response says so).
POST https://kairossignal.com/mcp
Tools: register_agent, list_products, purchase_data, check_balance, topup_credits
MCP also serves 5 paid analytics tools that debit credits per answered call, as each tool's description states. Paid per answered call: get_hodge_regime_sieve (1 credit), get_rmt_regime_state (1 credit), get_depin_slashing_audit (2 credits), get_shap_node_attribution (1.5 credits), get_listing_dilution_score (1 credit).
$1 = 1 credit · credits buy data products at their listed prices, delivered inline by purchase_data or POST /v1/credits/purchase · public GET browsing never debits credits, and credits do not unlock subscription tiers · One-time credit packs: $20 · $99 · current product prices
as_of, and a verify_yourself URL. Networks with no free public feed are explicitly named as such — we tell you what we don’t cover, not just what we do. “Try to break our numbers” is the pitch.