Noctel
NOCTELbeta
← back to terminal

HOW NOCTEL WORKS

Radical transparency is the entire product. Below is exactly what runs when, what it costs, and how to audit every signal you see.

The pipeline

Noctel is built on the Aeon framework inspired architecture for autonomous skill pipelines. The full execution log is public git history.

every 15 min → fetch-tweets, on-chain-monitor
every 1 hour → narrative-tracker, tweet-roundup
every 6 hours → digest, paper-pick
daily 7 AM UTC → morning-brief (composes “Today's Signal”)
daily 11 PM UTC → reflect, signal-verdict, cost-report

For live skill status, recent runs, and weekly cost see the public agent dashboard.

The source pool

Every signal is derived from a curated source list. X handles, long-form RSS feeds, and watched on-chain addresses. The full list is versioned in the public repo for anyone to audit.

We aim for balance: ~30% pro-crypto voices, ~30% skeptical voices, ~40% nuanced. The ratio is reviewed when the source list updates.

The models

Different tasks use different models, selected for cost vs quality fit, not loyalty:

Tweet extraction, classificationGPT-4o-mini / Haiku 4.5
Narrative clusteringHaiku 4.5
Today's Signal synthesisSonnet 4.6
Quality scoringHaiku 4.5

All LLM calls use prompt caching (90% discount on cached input). Every signal includes its token cost in the agent log.

The data layer

Noctel ingests from four streams: social posts, long-form essays, on-chain activity, and prediction markets. The what we read is public (every handle, feed, and wallet is versioned in the repo). The how (which providers and infrastructure we route through) is kept private. That plumbing is the only edge a public-good terminal has, and exposing it would let anyone clone the pipeline overnight.

Social postscurated handles → see repo
Long-form essayscurated feeds → see repo
On-chain eventswatched wallets → see repo
Prediction marketstracked markets → see repo

The cost

Running Noctel costs roughly $80-150/month at current traffic. The cost-report skill publishes a detailed breakdown weekly. We do not charge users. See the donate page if you want to help cover this.

The biases we acknowledge

  • English-only. Spanish, Chinese, and Korean crypto Twitter are huge and we miss them.
  • Curator bias. The source list reflects the curator's judgment of who's worth listening to.
  • Recency bias. The system prioritizes recent signal over historical context.
  • Crypto-Twitter bubble. We're aware crypto Twitter is itself an echo chamber, even when balanced internally.

How to audit a signal

  1. Click any signal's “agent log” link.
  2. See the skills that ran, candidates evaluated, why this won.
  3. All raw outputs live in the public GitHub repo under outputs/.
  4. Every prior signal is browseable as git history.
  5. If you find a signal that looks wrong, file an issue. We respond publicly.