AI Fund — Open-Source AI Hedge Fund for Claude Code

Hire your AI trading desk. Fire the ones that miss KPIs.

45 AI trading agents. 21 named personas (Arthur Hayes, Jim Simons, George Soros, Jesse Livermore, Warren Buffett, Peter Lynch…). 147 built-in MCP tools across Cube (31), CCXT (92), and Alpaca (24), plus 110 CCXT exchanges. Paper trading by default. MIT licensed. Runs on Claude Code.

<!-- GitHub Topics (set these in repo Settings > Topics): ai-trading, crypto-trading-bot, hedge-fund, ai-hedge-fund, trading-agents, mcp, claude-code, algorithmic-trading, market-making, arbitrage, quantitative-trading, risk-management, multi-exchange, defi, bitcoin, ethereum, crypto-fund -->

<p align="center"> <img src="docs/architecture-light.svg" alt="How AI Fund works — You talk to Claude Code, which orchestrates 45 trading agents across cube.exchange, Binance, Coinbase, Kraken, OKX, and 110 CCXT exchanges via MCP connectors" width="100%"> </p>

> /hire risk-manager
> /hire arthur-hayes
> /hire market-maker

> @arthur-hayes what's the macro thesis right now?
> scan all exchanges for BTC price differences
> the arbitrageur found a 15bps spread between cube.exchange and Binance — execute it
> risk-manager, approve this trade

---

What Is ai-fund?

45 autonomous trading agents inside Claude Code, with built-in MCP connectors and broad multi-exchange support via CCXT. Shared analysis libraries cover indicators, execution, portfolio analytics, and risk tooling.

21 are named personas — Arthur Hayes, Jim Simons, George Soros, Jesse Livermore, Stanley Druckenmiller, Warren Buffett. The other 24 are role-based: scalpers, market makers, risk managers, quants, arbitrageurs.

No config files. No YAML. You hire agents that fit your thesis and fire the ones that don't deliver. Each one carries its own personality, philosophy, and KPIs.

Count Snapshot (from this repo)

• 45 agents total in skills/ (_template excluded)
• 21 named personas + 24 role-based specialists
• 147 built-in MCP tools across active servers
  • Cube: 31 tools
  • CCXT: 92 tools
  • Alpaca: 24 tools
• 110 exchanges via CCXT (from the installed ccxt package in this repo)

How is this different from a grid bot?

You get a quant analyst that only trusts data, a risk manager that blocks trades when the sizing is wrong, and a market maker running Avellaneda-Stoikov across three venues.

The arbitrageur watches every connected exchange for mispricings and won't shut up about it. They argue with each other. The risk manager says no a lot.

More exchanges = smarter desk. Cross-exchange arb, smart order routing, multi-venue MM.

---

What You Can Do

| Strategy | Description | |----------|------------| | Cross-exchange arb | Spot price gaps, execute both legs | | Market making | Multi-venue quotes, Avellaneda-Stoikov | | Macro trading | DXY, yields, Fed policy via Hayes agent | | Stat arb / quant | Mean reversion, momentum, pairs | | Portfolio | Risk parity, Kelly sizing, drawdowns | | Execution algos | TWAP, VWAP, Iceberg routing |

---

How It Works

| Step | What Happens | |------|-------------| | 1. Connect | Connect any exchange via plugins. Paper mode by default. | | 2. Hire | Pick agents. KPIs get tracked. | | 3. Trade | Agents propose, risk manager approves. |

Paper trading is on by default. You have to opt in to live.

YOU (trader)
  │
  ├── /hire risk-manager          ← activate agents
  ├── /hire arbitrageur
  ├── /hire market-maker
  │
  ▼
CLAUDE CODE (AI runtime)
  │
  ├── Skills (45 SKILL.md files)  ← agent personas, strategies, KPIs
  │
  ├── Exchange Connectors (MCP)   ← connect any exchange
  │   ├── Cube (built-in)
  │   ├── Binance, Coinbase, Kraken, OKX...
  │   └── 110 via CCXT
  │
  ▼
YOUR EXCHANGES (paper or live)

Skills define what an agent thinks and does.

Connectors talk to exchanges.

The two layers don't know about each other. Add an exchange, don't touch agent code. Write an agent, don't touch exchange code.

---

Who Is This For?

| You are a... | ai-fund gives you... | |-------------|---------------------| | Crypto trader | 45 agents, natural language | | Quant | Backtest, stat tools, multi-exchange | | Fund operator | KPIs, hire/fire, risk controls | | Developer | MIT skill system, any exchange |

---

AI Contributor Path (for faster maintenance)

If you are modifying this repo with an AI coding agent, follow the maintainer guide:

• docs/ai-maintainer-path.md
• npm run repo:map for live architecture + risk hotspots
• npm run repo:map:json to generate .ai/repo-map.json and .ai/context-pack.md

This gives agents a machine-readable architecture map, dependency-cycle detection, and a prompt-ready context pack for safer edits. PRs also run the AI Repo Intelligence workflow to publish these artifacts automatically.

---

Quick Start

git clone https://github.com/cubexch/ai-fund
cd ai-fund
npm ci

Open Claude Code and connect your exchanges:

claude
> /setup

Hire your first agents:

> /hire risk-manager
> /hire arthur-hayes
> /hire jim-simons

Put them to work:

> @arthur-hayes what's the macro thesis? DXY is falling and the Fed paused.
> @jim-simons scan for statistical anomalies across BTC pairs on all exchanges
> @risk-manager size a long position given current portfolio

Cube CLI + Expanded Tool Surface

Recent Cube connector updates added a first-party cube CLI and expanded analysis/trade tooling.

cd connectors/cube/mcp-server
npm run login
npm run status
npm run cube -- --help

Use the command groups directly (cube account ..., cube market ..., cube order ..., cube risk ..., cube trade ...) or call MCP tools from Claude Code. Full reference: connectors/cube/README.md.

Advanced Tool Families (Cube)

• Portfolio/Risk intelligence: assess_portfolio_risk, simulate_stress_test
• Signal intelligence: detect_confluence, detect_bb_squeeze, get_market_microstructure
• Execution planning: plan_twap, simulate_market_impact
• Smart routing/discovery: search_assets, get_trending, execute_trade

These tools are designed to be chained (scan → validate → risk-check → execution plan → route).

---

Supported Exchanges (Counted)

Connect any exchange via MCP plugins. Each connector handles authentication differently — some use API keys, others use local auth. See connectors/README.md for setup details.

When handing an AI agent access to exchange credentials, think about security. See the API Key Security section below.

Crypto Exchanges

| | Cube | Coinbase | Binance | Kraken | OKX | |---|---|---|---|---|---| | Connector | Built-in MCP | Built-in (CCXT) | Built-in (CCXT) | CLI | okx-mcp | | Auth | Local (no keys) | Key+secret+pass | Key+secret | Key+secret | Key+secret+pass | | Spot | ✅ | ✅ | ✅ | ✅ | ✅ | | Perps | ✅ | Limited | ✅ | ✅ | ✅ | | Paper | ✅ | Sandbox | Testnet | ✅ | Demo |

Equities and Multi-Asset Platforms

| | Alpaca | Kraken | IBKR | |---|---|---|---| | Setup | Built-in MCP | CLI | CCXT/custom | | API keys | Key+secret | Key+secret | Portal auth | | Stocks | ✅ No-fee | Tokenized | ✅ | | Crypto | ✅ | ✅ | ❌ | | Options | ❌ | ❌ | ✅ | | Paper | ✅ | ✅ | ✅ |

API Key Security — Why This Matters With AI Agents

AI agents can read files, call tools, log output, and spawn processes. Your API key in a config file? The agent can see it.

| Risk | Mitigation | |------|-----------| | Keys in config | Use read-only keys. Disable withdrawal. | | Keys in env vars | Use subaccounts with limited funds. | | Keys in logs | Avoid verbose mode. Review MCP code. | | Keys in transcripts | Scrub before sharing sessions. | | No rotation | Rotate regularly. IP whitelist. | | Withdrawal enabled | Always disable. Use subaccounts. |

Some connectors (like Cube's built-in MCP) use local auth with no API keys. Others require key+secret in config files. Choose connectors that match your security requirements.

110 additional exchanges work via npm i -g ccxt-mcp — anything CCXT supports in this repo's pinned CCXT version.

More venues = more strategies. Cross-exchange arb doesn't work with one exchange.

See connectors/README.md for setup details.

---

45 AI Trading Agents — The Full Roster

Named Personas

21 agents modeled after real traders. The philosophy isn't just flavor text — it changes how they read markets and size positions.

| Persona | Philosophy | Style | |---------|-----------|-------| | Arthur Hayes | Macro-to-crypto. DXY, real yields, liquidity cycles. | Leveraged macro conviction | | George Soros | Reflexivity theory. Attack regime breaks. Boom-bust cycles. | Thesis-driven, concentrated | | Stanley Druckenmiller | Go for the jugular. Concentrated macro bets when conviction is high. | High-conviction sizing | | Paul Tudor Jones | Risk management IS the strategy. 200-day MA. 5:1 R:R minimum. | Trend following, risk-first | | Ray Dalio | All-weather portfolio. Risk parity. 15 uncorrelated bets. | Balanced allocation | | **[Jim Simons](/skills/jim-sim