kubera

An autonomous multi-agent quant research firm powered by Claude Code

Researchers generate signals. Reviewers filter. Tech backtests via Flint. Strategies auto-promote through paper trading to live. The only human gate: real capital deployment.

How It Works

Kubera is an autonomous research pipeline that runs as a background daemon. It spawns Claude Code sessions as specialized agents — researchers, reviewers, and technicians — that collaborate through a shared ledger to discover, validate, and deploy quantitative trading strategies on crypto perpetuals.

Key Design Decisions

• Researchers write real code, not pseudocode. Each signal includes a working strategy.py that the reviewer re-runs. No gap between "what was proposed" and "what was tested."
• Tech is zero-token. Backtests run as deterministic Python through the Flint bridge — no LLM in the loop. Cost scales with researcher + reviewer sessions only.
• Walk-forward replaces proxies. Real N-fold walk-forward validation, not oos_sharpe * 0.5.
• Risk-free rate adjustment. OOS Sharpe subtracts the risk-free rate so long-biased strategies can't coast on rate income.

Quick Start

Prerequisites

• Python 3.11+
• Claude Code CLI installed and authenticated via Max subscription
• Flint cloned as a sibling directory

Architecture

Ledger Schema

The SQLite ledger (firm/ledger.db) tracks: