The agent feed, MCP methodology

What it is

One MCP call returns one normalized market-state object. The agent feed covers seven assets, BTC, ETH, SOL, BNB, XRP, DOGE, HYPE, across Binance, Bybit, OKX and Hyperliquid, published under the registry name ai.markettrace/agent-feed and hosted at https://api.markettrace.ai/mcp.

The feed reports what it can measure and self-declares how much to trust each number. It never returns a trade recommendation; it reports history and coverage-honest facts. It reports history, not predictions.

Data sources

The feed reuses the same cross-exchange pipeline that powers the site , all-Rust, zero Python daemons. Each metric carries its own provenance and links to a deeper per-metric methodology:

• Funding. An OI-informed cross-venue rate plus a multi-year percentile over a consistent, deep venue set. Funding methodology →
• Open interest. A 4-venue native-USD sum: Bybit openInterestValue, OKX oiUsd, Hyperliquid coins × markPx, Binance coins × mark.
• Volume. A 30-day consolidated tape.
• CVD & order-book imbalance. Cross-venue taker flow and flow-weighted OBI. Positioning methodology →
• Liquidations. Three venues, Binance, Bybit, OKX; Hyperliquid is deliberately skipped. Liquidations methodology →
• Basis. Perp-vs-spot in basis points; positive means the perp trades above the spot index, where the index is a multi-exchange spot composite.
• Price / OHLCV. Consolidated cross-exchange candles.

Open interest sums the four venues in native USD, Bybit openInterestValue, OKX oiUsd, Hyperliquid coins × markPx, Binance coins × mark. oi.rel compares the current value against its trailing hourly median (1d / 1w / 1m), and oi.chg_1h_pct is the change over the previous complete hour. The baseline joins on the live venue set, so legacy Binance-only rows drop out of the median, an honest re-accrual rather than a silent blend of old and new coverage.

volume.usd_24h is the consolidated 24-hour traded notional across venues; volume.rel expresses that against its trailing hourly median, as a multiplier.

price.last is the consolidated mid. The chg_*_pct fields are returns versus a tolerance-matched baseline; atr_1h_pct is ATR(14) on hourly bars, as a percent of price; and rv_24h_pct is realized volatility annualized from minute returns.

Coverage honesty

Every metric carries a coverage entry. A thin or young metric is answered honestly, with its depth disclosed, and is never faked to look deep.

coverage: { venues, window_days, n_samples, partial, reason }
reason ∈ { accruing | unavailable | degraded | stale }

age_seconds is the worst-case age across the live sources that fed the non-null fields, so a single stale sub-feed cannot hide behind fresher ones. A feed beacon on every response , feed.version and feed.tools, lets an agent detect a stale, cached tool catalog and re-read the schema.

Conditional outcomes

The flagship tool measures base rates of forward returns after a caller-stated condition. It replaces folklore, “high funding means a squeeze”, with the base rate drawn from the feed's own data. The mechanics are deliberately conservative:

• As-of percentile. At each historical hour t, a condition stated as a percentile is ranked only against data at or before t, no look-ahead. A structural floor applies: an as-of rank exists only once its series holds 12 samples, because a rank among two points is degenerate. Raw-value thresholds carry no such floor.
• Matching. A match is any hour where every stated condition field holds at once, the fields AND together.
• Forward returns. Per horizon (default 4h / 24h / 72h):

r_h(t)  = close(t+h) / close(t) − 1            // forward return at horizon h
maxDD_h = min( low(t+1 … t+h) ) / close(t) − 1  // worst hourly low in (t, t+h]

• Overlap collapse. Overlapping forward windows are not independent samples, so matches are thinned to at least one horizon apart, yielding n_effective; the raw n_matches is reported alongside it. Every statistic is computed on n_effective.
• Cluster warning. The response reports the share of effective matches falling inside the densest 14-day window, a guard against one squeeze episode masquerading as N independent episodes.
• The history_silent verdict. Below 12 effective matches on every horizon, the statistics are null and the verdict reads history_silent. This is a first-class answer, “the history is silent; you would be trading on belief, so size accordingly”, not an error.
• Price series. Returns are computed on Binance 1-hour closes (coverage.price_src), by design: no consolidated or venue-stitched series inside a single backtest.
• Archive-feature (v2) conditions. Conditions on obi_skew, oi_chg_1h_pct, taker_buy_ratio or basis_bps intersect the funding window with the young 15-minute archive; the response window reports that intersection honestly.

State history

A 15-minute archive stores the full served market-state per asset. get_state_history returns parallel time-series arrays of any numeric dotted field, funding.percentile, oi.usd, obi.skew, downsampled to max_points. The stride used is reported, and the newest row is never sampled away. The archive is young, it was born 2026-07-03, and grows forward, so thin answers are honest, not broken.

Tools

• get_market_state, one normalized market-state object for an asset.
• get_funding_percentile, funding rate with its multi-year percentile.
• get_liquidations_recent, recent liquidations across the three covered venues.
• get_ohlcv, consolidated cross-exchange candles.
• get_conditional_outcomes, base-rate forward returns after a stated condition.
• get_state_history, time-series arrays from the 15-minute archive.

For how to connect an agent, see /agents.

Limitations

• Binance-only price in the backtest. The conditional-outcomes series uses Binance 1-hour closes; the delta versus a consolidated price is small at the 4h-and-longer horizons it reports.
• The state archive is young. It grows forward from 2026-07-03, so archive-feature conditions and get_state_history return short windows until it fills.
• Inherited coverage caveats.Each underlying metric's own caveats carry through, weight any reading by its coverage entry.
• Read-only, descriptive. The feed is read-only and never a trade recommendation. It reports history, not predictions.

Versioning

Methodology version v1.4.1 · updated 2026-07-04. Material changes (new sources, formula tweaks, threshold changes) bump the version and update dateModified in the structured data above.

v1.4.1 (2026-07-04): first published methodology for the MCP agent feed, the coverage-honesty model, conditional-outcomes statistics, and the 15-minute state archive.