How it works
Glint is a small set of services that turn raw market activity into ranked signals and push them to your screen. This page is the honest mid-depth tour: which pieces exist, what they do, and which third parties power them.
The pipeline in one sentence
Per-chain indexers stream market data into a central Postgres, an engine ticks every minute to score new tokens against per-chain rules, a tracker keeps updating the peak gain for every recent call, an AI service writes a one-line explanation per signal, and a Telegram bot pushes the results to subscribers.
Per-chain indexers
The scoring engine
The engine runs every 60 seconds. On each tick it asks each chain for the list of tokens that have been active in the last hour, scores each one through the chain's dedicated scorer, and inserts a signal row when the score crosses the threshold and the per-token cooldown has passed.
Each chain has its own scoring file because the inputs are different (DEX swaps vs perp trades vs pump.fun snapshots), but the output is always the same shape: a number from 0 to 100, split into momentum (0-50) and whale flow (0-50), plus a raw payload of the intermediate numbers the AI explanation uses to write its sentence.
The per-chain rules are spelled out on the Chains page. The scorer is tuned against historical outcomes; the pump.fun scorer was re-tuned after the first 381 calls based on which mcap and ATH-proximity bands actually delivered.
Peak tracker
A separate tracker runs every 60 seconds, scans every signal from the last 30 days, and updates the high-water mark per call. If the current price beats the recorded peak (or beats the at-call price on a never-pumped signal), the new peak is written.
That is how the "Peak" column on cards and the "PEAK GAIN" stat on token pages stay honest over time. The peak never goes down. Even a round-tripped runner reads as a winner here, because the question the column answers is "what is the most this call ever gave you".
AI explanations
Each signal carries a one or two-sentence explanation in plain language ("9x volume vs the 24h baseline with three smart-money wallets buying, late-stage pump risk in play"). That sentence is generated by Anthropic's Claude (Sonnet 4.6) with the raw scoring payload as input.
The prompt is short and instructs the model to lead with what is happening on chain, mention direction when present (Hyperliquid), and avoid hype words. The same model is used across all chains so the tone is consistent.
The web dashboard
Next.js 15, server-rendered. Pages call the engine API for fresh data on every request, with a short revalidation window so multiple loads in the same minute share cached enrichment. The signal cards, filters, sort options, and call history are all client components reading from server-rendered JSON.
The Telegram bot
A small Telegraf-based bot subscribes to two streams in the engine database: new threshold signals (gated by each subscriber's /threshold) and watchlist signals (every new signal for tokens on a subscriber's /watching list). Polls every 30 seconds. Both streams are deduped per subscriber so the same signal never fires twice.
What is not in this pipeline
Glint does not hold custody of anything, does not route trades, does not run a paid market data feed. Everything you see is built on public data and public APIs. The revenue model on the roadmap is the VIP tier gated by holding $GLINT, which buys you access to a lower alert threshold, faster polling, smart-money-only filtering, and priority push.