BPE Market Intelligence

A skill for AI agents that need to reason about Bitcoin markets across venues. Pairs naturally with single-exchange MCP servers (Bybit, Coinbase, Gemini, Kraken, OKX) — those handle execution on one venue; BPE handles cross-venue intelligence.

When to use this skill

Reach for BPE tools when the agent's task involves any of:

• "What's BTC doing right now?" → call get_market_briefing first.

One call replaces dozens of raw-data calls plus the synthesis work.

• Spread / arbitrage / basis questions → get_funding_skew for

perp-perp; the briefing endpoint surfaces consolidated-vs-venue divergence.

• Directional questions ("is the market bullish?", "should I expect

mean-reversion?") → get_ml_signal returns calibrated probability, not just direction.

• Reconciling a position against the broader market → if the user

is on a single exchange and asks why their P&L looks the way it does, cross-check the venue's price against get_consolidated_price.

• Anomaly detection → the briefing's anomaly hint flags extreme

funding spreads or high ML confidence regimes worth investigating.

When NOT to use this skill

• Order routing, fills, balances, withdrawals → use the user's

exchange MCP (Bybit, Coinbase, etc.). BPE is read-only and has no execution surface.

• Exchange-specific account state ("what's my Bybit balance?") →

exchange MCP, not BPE. We don't have visibility into user accounts.

• Wallet operations, signing, on-chain → not this skill.

Recommended call patterns

Pattern 1 — orient before drilling down

For open-ended market questions, always call `get_market_briefing` first. It costs ~1 rate-limit token and ~50 tokens of context, and gives the agent enough situational awareness to decide what (if anything) to drill into next. Avoid making the agent assemble a briefing from individual get_consolidated_price / get_funding_skew / get_ml_signal calls — that's slower, more expensive, and what the briefing endpoint exists to prevent.

Pattern 2 — confidence-gated decision-making

get_ml_signal returns calibrated probabilities. Confidence is on a 0–1 scale where:

• < 0.55 — treat as noise. The model has a slight edge but not

enough to justify a thesis.

• 0.55–0.70 — directionally interesting. Worth combining with other

signals (funding skew, sentiment, regime) before acting.

• > 0.70 — high-conviction. Surfaced as an anomaly in the briefing.
• > 0.85 — rare; the model is unusually confident. Worth flagging

to the user explicitly.

Don't treat confidence as binary. Don't translate "up" / "down" into trade signals without considering the magnitude.

Pattern 3 — companion to an exchange MCP

When the user has both an exchange MCP and BPE installed:

• Use the exchange MCP for: order placement, account state, fills,

balance queries, withdrawals.

• Use BPE for: pre-trade context (is the market right for this

trade?), post-trade reconciliation (did my venue's price match consensus?), strategy ideation (where's funding cheapest?).

• Never confuse the two. A BPE call that reads get_consolidated_price

doesn't reflect the price the user can actually execute at on their venue. Use the exchange MCP's order-book or quote tools for that.

Tool reference

All tools return text content blocks designed for direct consumption by the LLM. No JSON parsing required.

Available time horizons

get_ml_signal accepts 30s, 60s, 5m. Available horizons may evolve over time — if the agent receives an "available horizons" error, it should retry with one of the listed alternatives.

Limitations to surface to the user

• Read-only. BPE cannot place orders, manage positions, or move

funds. If the user asks the agent to act on a signal, route through their exchange MCP.

• BTC-only in v0.1. ETH, SOL, and other assets are out of scope.

If the user asks about another asset, say so plainly.

• ML signal is a probability, not a guarantee. The model is well-

calibrated but markets remain noisy. Frame predictions as "the model thinks X is more likely than Y", not "X will happen".

Background

BPE (Bitcoin Pricing Engine) consolidates 36 exchanges into a single normalised market-data feed with sub-100µs internal latency, plus derived intelligence layers: a calibrated gradient-boosted ML model trained on a broad engineered feature set, sentiment indicators (Fear & Greed, news, mempool stress), and cross-venue funding / basis analysis. Running in production for institutional users; the MCP server exposes this as native agent tools.