The Core Architecture (Perps Inception Loop)
Data Ingestion Layer
Real-time and historical market data flows into the agent :Price & orderbook from Phoenix
Funding rates, open interest, liquidations
On-chain flows, sentiment, and macro signals
Phoenix’s fast settlement (~0.5 seconds) and deep liquidity make it ideal for agentic trading.
Because of its strong financial reasoning and long context window, it can maintain memory across many trade cycles and evolve its own style over time. Claude Fable 5 Brain
Fable 5 receives the full context and is given complete strategic free will.
It performs deep multi-layered analysis: Technical structure & key levels
Narrative momentum & regime detection
Risk assessment & expected value calculations
Psychological/market sentiment undercurrents
Decision & Execution Layer (via Vulcan CLI MCP)
Fable 5 decides on a trade (direction, leverage, size, entry/exit rules).
It then uses Vulcan — Phoenix’s official CLI built specifically for AI agents. Vulcan exposes tools via a local Model Context Protocol (MCP) server.
Claude Code / Claude agents can directly call Vulcan commands to place orders, manage positions, adjust margin, etc.
Everything runs on-chain through Phoenix’s orderbook or AMM mechanisms.
Feedback & Evolution Loop
Every trade (win, loss, or flat) is fully logged with reasoning.
These logs are fed back into Fable 5 in the next cycle.
Over time, the agent develops its own “personality,” risk appetite, and edge — just like my Infinite Backrooms project, but now applied to perpetual markets.
Why This Combination Works Extremely Well
Phoenix → High-performance on-chain perps with excellent developer tools (Vulcan MCP was practically made for this).
Claude Fable 5 → Top-tier reasoning on trading analysis, expected value, and autonomous agent tasks.
Solana → Sub-second execution low fees = perfect environment for a live, always-on agent.