How Senpi Trades: The Five-Phase Engine
The shared infrastructure every strategy runs on. Each trade flows through five phases — click any to jump to its capabilities.
Find what's worth trading from 500+ perps.
Score it. Gate it. Confirm conviction.
Open the position with optimal fees.
Defend, then ride. Two phases of risk.
Close cleanly. Audit the trade. Learn.
Finding what's worth trading
Senpi runs four parallel signal layers to filter 500+ Hyperliquid perps down to a handful of opportunities per cycle.
Scores every Hyperliquid perpetual 0–400 across four independent dimensions. Runs every tick, surfaces only the highest-conviction setups to the strategy layer.
Streaming data pipeline from Hyperliquid's top-trader leaderboard. Rank changes, position-level delta PnL, momentum events, per-market concentration — all from real trader behavior, not synthetic signals.
Tracks Smart Money market-concentration rank velocity across all 500+ perps. The signal is how fast an asset is climbing — not where it currently sits. Flags assets early, before the cohort positions and before price moves.
Aggregates top-trader notional exposure across the full leaderboard cohort, per asset. Where Emerging Movers is timing, Whale Index is conviction — how much capital the smartest cohort has actually deployed.
Scoring, gating, and conviction
A surfaced opportunity isn't a trade — yet. Senpi runs each candidate through configurable confluence gates and risk guard rails before any capital is committed.
Each strategy declares its own gate stack — technical, hyperfeed, structure, funding. A trade only fires when every required gate clears. Strict-gate strategies (Cheetah, Condor) trade rarely but high-conviction; relaxed gates (Vulture) trade more often.
Runtime v3.0 reads risk.guard_rails from every strategy config. Position size limits, max leverage, concurrent position counts, asset exclusions — all enforced at the runtime layer before MCP is ever called.
Opening the position
When a trade fires, the runtime owns it. Fee-optimized limit orders are placed natively. Every step is logged. Skills never call MCP directly.
In-process producer daemon with direct HTTPS to MCP. Native FEE_OPTIMIZED_LIMIT entries and exits. Full trade-chain DB telemetry. Clean lifecycle hooks for skill authors. The runtime is the only caller — skills never talk to MCP.
Decides per-trade whether to use FEE_OPTIMIZED_LIMIT (maker rebate, passive fill) versus MARKET orders, based on urgency, spread, and position context. Maker rebates on Hyperliquid can flip the P&L on marginal trades. Compounds at volume.
Cutting losers, riding winners
The DSL Exit Engine is the single most impactful component for long-run P&L. Two phases: defense first, then trail. The position only moves forward.
Closing cleanly, learning from every trade
Exits run through the same fee-optimized rails as entries. Every closed trade lands in the audit log — searchable, queryable, and feeding the next iteration of strategy improvement.
Full chronological record of every decision a strategy makes — signals fired, orders placed, position changes, exits triggered. Queryable by time, asset, action type. The audit log is how strategy authors find bugs and how users build trust.
Every capability is exposed as a typed MCP tool. Structured API surface covering the full lifecycle — discovery through execution through audit. Browse the surface below or use it in the Dojo.
Browse the API surface
Every capability is exposed as a typed tool. Click a category to see its tools — names and descriptions inline.