Moonlytics.io
Use cases / AI developers

AI developers

Retrieval and agent tooling over digital asset data. The hard part is not access — it is making sure the model cannot confidently answer a question the data does not support.

Three failure modes worth designing against

A language model given crypto data will produce a fluent answer regardless of whether the data justifies one. The failures are predictable, which means they are preventable at the data layer.

1. Ambiguous identity

The most common and the least visible. A user names a ticker; the ticker belongs to several assets; the model answers about one of them without mentioning the ambiguity.

ambiguity is a real state, not an error
user: "how is HYPER doing?"

# The registry has three assets under that ticker:
#   HyperChainX · Hyperlane · Hyperpigmentation
#
# A resolver that returns one of them makes the model
# confidently wrong. A resolver that returns three lets
# the model ask.

Because the layer resolves to candidates rather than to a winner, "which one do you mean?" is an answer the model can actually give. The resolver behaves this way, and so would a tool built on it.

2. Implied currency

A number without a timestamp reads as "now". The fastest cadence here is hourly, so a price is up to sixty minutes old — a fact that must be in the payload, not in a footnote a model never sees. Measured cadence per category.

3. Inferred causation

Given sentiment and price in the same context, a model will narrate one as causing the other. The layer's countermeasure is documentation: every category states its limits in the same eight-question structure, and that text is what a tool description should carry so the boundary is in the model's context before it answers.

What makes this data retrieval-friendly

  • Consistent structure. One asset identity, one schema, one time base across nine categories.
  • Self-describing values. Every observation carries its asset, its timestamp and its unit.
  • Documented limits per category, in prose a model can be given verbatim.
  • Prose already written for machines. The news feed carries self-contained summaries that name their entities and explain jargon inline — written that way for human readers, and unusually good retrieval chunks as a result.
  • Narrative snapshots with source counts — evidence attached to a claim, which is what lets an agent say "eighteen sources" instead of "everyone is talking about".

Current access

No agent interface exists yet

There is no MCP server, no agent tool API and no public REST endpoint today. The intended tool surface and the reasoning behind its shape are set out under AI agents; nothing there is installable, and it is labelled as planned throughout.

What this data will not support

  • Trading signals or recommendations — the layer carries descriptive data only.
  • Sub-hour reaction; the fastest cadence is hourly.
  • On-chain state, balances or transaction tracing.
  • Claims about future price. No dataset here supports one.

Coverage summary for context windows

7,461 canonical assets; 6.7 million hourly market, social and sentiment observations since November 2025; 1,000 assets per hourly snapshot; daily ecosystem series for 69 assets; hourly news window and hourly narrative snapshots. Fastest cadence hourly. Measured 2026-08-31.

Related