claudefiautonomous claude agents across defi

self-improving ai that learns from every trade

get started

view on github

self-improving

learns from every trade. wins become patterns, losses become warnings. skills evolve over time.

4 trading domains

dlmm liquidity provision, perpetual futures, spot memecoins, and prediction markets.

risk controls

multiple layers of validation - position limits, drawdown guards, confidence thresholds.

claude agent sdk

multi-turn tool conversations for complex decision-making with full context.

mcp server

custom model context protocol server exposing domain-specific trading tools.

hooks system

event-driven middleware for validation, logging, and custom guard rails.

how it works

trade → outcome → analysis → skill → better future trades
           ↑                    │
           └────────────────────┘
              feedback loop

claudefi runs a continuous 30-minute decision cycle called the ralph loop:

observe - fetch live market data across all domains
think - claude analyzes opportunities with learned skills
act - execute validated decisions through hooks
learn - generate skills from outcomes

quick start

bash

bunx claudefi

that's it. paper trading is on by default.

trading domains

domain	platform	strategy
dlmm	meteora	concentrated liquidity provision
perps	hyperliquid	rsi-based futures trading
spot	jupiter	memecoin trend trading
polymarket	gamma api	probability edge detection

cost estimation

component	per cycle	daily (48 cycles)
decision making	~$0.15	~$7.20
judge evaluation	~$0.08	~$3.84
skill generation	~$0.04	~$1.92
total	~$0.27	~$13