Classifies narrative propagation in crypto markets — which explanatory framework is spreading, how fast, and at what confidence — weighted against traditional market dynamics and prediction-market probability. Delivered as signal packs with narration.
AGORA VALLEYS
Autonomous AI market intelligence
A market sensor built on primary sources.
PHEME reads four layers of the market continuously — news, prediction markets, macro and crypto positioning — and applies semantic and event classification to each cycle. It writes one signal object every fifteen minutes, unattended, with no editorial layer. Output is delivered as signal packs with narration attached: the classification, the scoring, and the sentence that explains it.
Semantic and event classification. Scored, weighted, and packed each cycle.
Event axis
What happened
Market axis
What the market is doing
15 min
Every cycle, unattended
8
Skills live now on the MCP server
5
Live channels, one signal object
<100ms
Pre-computed response, no LLM in the middle
01
What it produces
One reading, four layers deep.
Every cycle writes the same object: the classified news read, the macro backdrop, crypto capital positioning, and a single verdict across all three. Each layer is callable on its own, or together as one pack.
// regime — what happened, what the market is doingevent_regime"NEUTRAL"market_regime"NEUTRAL"polymarket 0.945
fear_greed 69 "Greed"
Gold bullish while BTC decouples from it. Capital concentrated in BTC, stablecoins flat. Greed at 69 against a soft tape.
All four layers return in one call, pre-computed, with narration attached. pheme_market_brief returns all of it as a single narratable string.
02
Classification model
Two reads per cycle: data and narration.
The event axis reads what happened. The market axis reads what the market is doing, derived from price, macro and flow. Precision comes from keeping them separate.
A cycle where the two axes disagreed — 2026-07-13, 13:57 UTC
The news said
ETF_CATALYST
intensity 0.68 · quality 0.82
The market was
RISK_OFF
price, macro and flow
A single-classification feed reports the bullish catalyst only. The market state was the opposite, and is reported separately.
Data — what PHEME writes
event_regime"ETF_CATALYST"
market_regime"RISK_OFF"
intensity0.68
signal_quality0.82
Narration — a product built on it
PREDICTABILITY — PUBLISHED READ
event_context"ETF_CATALYST"
market_regime"RISK_OFF"
event_intensity"ELEVATED"
read_confidence"HIGH"
"Market conditions read RISK_OFF with an ETF_CATALYST event overlay."
361
Cycles where the two axes disagreed
54
Of those, market read RISK_OFF
67 days
Measured window
03
Predictability — predictor scoreboard
The bar a healthy predictor has to clear.
Target thresholds for the downstream predictor, scored against a persistence baseline. Accuracy alone is not the test — anticipating regime transitions is.
1-hour horizon — target (n=275)
model_accuracy0.81
persistence_baseline0.76
delta+0.05
transitions_observed13
correctly_anticipated8
transition_recall0.62
distribution_recall0.36
19 of 53 realized
4-hour horizon — target (n=261)
model_accuracy0.74
persistence_baseline0.70
delta+0.04
transitions_observed35
correctly_anticipated19
transition_recall0.54
distribution_recall0.31
24 of 77 realized
Crowd fade — target rate by state
State
1h
4h
EXTREME_FEAR
0.76
0.82
FEAR
0.60
0.63
GREED
0.58
0.55
Reading it
A small accuracy margin over persistence is still real when it is won on transitions. Extremes fade easily; the meaningful test is that no single crowd state sits below random across both horizons.
04
Distribution channels
MCP-native signal delivery across five live channels.
Agent AI infrastructure
MCP server
8 skills, pre-computed and open. Every other channel consumes this server.
The same brief ships deterministically or with an AI synthesis layer on top. Long-form editorial runs as a separate pipeline.
Telegram — no LLM
🔴 Market Brief — Sept 08, 2026
BTC $79,258 ▼1.1%▲0.7% 7d
ETH ▲0.8% · SOL ▲0.5%
Mkt Cap $2,694B · Dom 59.2%
🔴 Risk-Off · Bullish · Low intensity Agreement Medium · Priced in Low
⚪️ Ripple CEO: Dutch central bank moved $11B in gold
⚪️ Japan Q2 GDP revised up, BoJ hike expectations
⚪️ Token buybacks hit a record $638 million + 5 more
Prices, market state and scored headlines. Fully deterministic — no model in the path.
BTC $66,330 ▲1.0%▲5.0% 7d
ETH ▲5.6% · SOL ▲11.3%
Mkt Cap $2,268B · Dom 58.6%
🟠 Cautious · Moderate intensity Agreement High · Priced in Medium
🤖 PHEME ANALYSISBOJ rate hike to 1% is the dominant macro signal, historically a headwind for risk assets including crypto; institutional Bitcoin accumulation adds noise but does not override tightening pressure.
⚪️ DeepSeek raises $7B+, valuation exceeds $50B + 7 more
Identical structure, with a synthesis block reading the layers together.
claude mcp add--transport http quanturama https://mcp.quanturama.com/mcp
Every skill is callable immediately — no key, no quota, no onboarding step.
Sample tool call
{
"tool": "call_skill",
"arguments": {
"skill_name": "pheme_regime",
"output_format": "text"
}
}
Sample response
{
"skill": "pheme_regime",
"result": "BEARISH CONFLUENCE",
"confidence": 0.81,
"layers": { "sentiment": "FEAR",
"funding_rate": "NEGATIVE",
"polymarket": "RISK_OFF",
"fear_greed": 22 }
}
Skill
What it answers
pheme_regime
What kind of event is live, and what the market is doing
pheme_market_brief
The whole market read in one call
pheme_macro_context
Gold, DXY, equities — and how BTC correlates to each
pheme_crypto_positioning
Market cap, dominance, where capital is rotating
pheme_macro_analysis
The cross-layer verdict, conclusion first
pheme_top_influencers
Which accounts are driving the narrative
07
Agent AI marketplace
Listed on Agora Valleys. Callable as a demo.
PHEME is published on Agora Valleys as callable skills under the marketplace's own signal_packs category. Each skill has a live page where the response can be triggered and read before any integration work.
Try it
Run a skill in the browser
Open a PHEME skill on the marketplace and call it. The response returned is the live one.
No key
Immediately callable
Every listed skill is live and free to call. No account, quota or onboarding step.
An independent product sells market context to professionals over REST and MCP, built as a client of PHEME's public MCP server — with its own vocabulary layer over PHEME's read.
Production consumer
Live quant system
A trading system consumes PHEME through the same public MCP server and adjusts position sizing on the classification.
4
Agent platforms live
8
Callable skills
2
Products built on the MCP server
11
Contact and references
Autonomous AI market intelligence, callable today.
A product of Quanturama.
claude mcp add--transport http quanturama https://mcp.quanturama.com/mcp