Skip to content

Chris0x88/hyperliquid-agent

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

155 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HyperLiquid Agent

HyperLiquid Agent

Your AI trading co-pilot for Hyperliquid

Open source. No fees. No telemetry. Your keys, your rules.

Python License Claude Code Telegram OpenClaw HyperLiquid


What Is This?

An open-source toolkit that lets you run an AI trading agent on Hyperliquid — the onchain perpetual futures exchange. It combines a tick-based trading daemon with Claude Code integration, so your AI agent can monitor positions, execute trades, and improve its own codebase over time.

What makes it different: This isn't a black-box bot you deploy and pray. It's designed to work with you. You bring the thesis and market knowledge. The agent brings discipline, execution speed, and 24/7 monitoring. Three interfaces, each doing what it's best at:

  • Claude Code (phone or desktop) — your AI trading brain. Deep analysis, code improvement, trade execution, strategy discussions
  • Telegram bot — instant dashboard. Portfolio, charts, prices, orders. Fixed Python code, zero AI credits
  • OpenClaw (optional) — AI conversation via Telegram DM for quick questions and market chat

The agent trades autonomously within the boundaries you set, and gets smarter through a recursive research loop.

Currently trading: Bitcoin Power Law rebalancing (vault), Brent Oil directional (main account). The system supports any Hyperliquid market including trade.xyz perps (oil, gold, equities, etc.).


5-Minute Setup

You need: Python 3.10+, a HyperLiquid account, and Claude Code.

Step 1: Clone and Install

git clone https://github.com/Chris0x88/hyperliquid-agent.git
cd hyperliquid-agent && pip install -e .

Step 2: Create an API Wallet (Critical)

Never give this tool your main private key. Use an API wallet instead.

HyperLiquid API wallets (agent wallets) can trade but cannot withdraw funds. If your key leaks, attackers can't steal your money — they can only make trades, and you can revoke the key instantly.

  1. Log in at app.hyperliquid.xyz
  2. Go to Portfolio > API Wallets > Generate
  3. Name it (e.g., "agent-bot")
  4. Copy the private key — you only see it once
# Import your API wallet key (encrypted storage)
hl keys import --backend ows

For a deeper explanation of API wallet security, see docs/SECURITY.md.

Step 3: Start the Daemon

# Watch mode — monitors only, no trading (start here)
hl daemon start --tier watch

# When ready for real trading:
hl daemon start --tier rebalance --mainnet

Step 4: Connect Telegram (Recommended)

Get instant trade alerts and send commands from your phone.

  1. Create a bot via @BotFather on Telegram
  2. Store the token securely (macOS):
    security add-generic-password -s hl-agent-telegram -a bot_token -w "YOUR_TOKEN" -U
  3. Message your bot, then run:
    python3 -c "
    import requests
    token = 'YOUR_TOKEN'
    r = requests.get(f'https://api.telegram.org/bot{token}/getUpdates')
    chat_id = r.json()['result'][0]['message']['chat']['id']
    print(f'Chat ID: {chat_id}')
    "
  4. Store the chat ID:
    security add-generic-password -s hl-agent-telegram -a chat_id -w "YOUR_CHAT_ID" -U

Now start the Telegram bot:

hl telegram start

Your bot responds to /status, /chart oil 72, /watchlist, /price, /pnl, /orders, /powerlaw. All instant, all free — fixed Python code hitting the HyperLiquid API directly. Zero AI credits.

Step 5: Set Up Claude Code as Your AI Trader

This is the real power. Open Claude Code (phone, desktop, or web) and point it at this repo. Claude can:

  • Make autonomous trading decisions
  • Run on a schedule (hourly check-ins via scheduled tasks)
  • Execute trades, manage positions, adjust leverage
  • Improve the codebase as it learns from each trade
  • Send you alerts via Telegram when something happens

Claude Code is the brain. Telegram is the dashboard. They're separate interfaces — Claude Code doesn't run through Telegram.

Step 6: Add OpenClaw (Optional — AI Chat via Telegram)

If you want AI conversation in Telegram (not just fixed commands), set up OpenClaw with a separate bot. This gives you a Telegram DM where you can ask questions like "what's the oil thesis?" and get AI responses.

See docs/openclaw-setup/ for the agent prompt and configuration.

Note: OpenClaw and the commands bot run as separate Telegram DMs, not in a group. Multi-bot groups don't work cleanly with Telegram's current architecture (see docs/TELEGRAM_GROUP_SETUP.md for why).


How It Works

The Daemon

A tick-based loop (inspired by Hummingbot's clock architecture) that runs registered iterators in dependency order:

Every tick (default 60s):
  Connector  → fetch prices, positions, balances
  Liquidity  → detect time-of-day regime (weekend/after-hours = danger)
  Risk       → check drawdown, circuit breakers, gate state
  Guard      → trailing stops, two-phase profit protection
  Rebalancer → run strategies, convert decisions to orders
  Profit Lock → sweep 25% of realized profits (capital protection)
  Journal    → log tick snapshot for analysis
  Telegram   → send alerts for trades, gate changes, P&L

Three Tiers

Tier What It Does Start Here?
watch Monitors only. No trading. Sends alerts. Yes
rebalance Auto-rebalances + guards positions with trailing stops. When you trust it
opportunistic All above + scans for opportunities via Radar/Pulse. Advanced

Liquidity-Aware Risk

The daemon knows when markets are thin and adjusts automatically:

Regime When Size Multiplier Stop Width
Normal Weekday, US/EU hours 1.0x 1.0x
Low After-hours (22:00-06:00 UTC) 0.6x 1.3x
Weekend Saturday/Sunday 0.4x 1.5x
Dangerous Weekend + after-hours 0.25x 2.0x

This matters. Low-liquidity stop hunts are how retail traders get wiped. The daemon reduces exposure automatically.

Profit Locking

The ProfitLockIterator sweeps 25% of realized profits by partially closing profitable positions. This protects your capital base — profits are "locked" even if the trade reverses.


Supported Markets

The agent trades on both native Hyperliquid perps and trade.xyz builder-deployed perps:

Market Type Examples Coin Format
Native perps BTC, ETH, SOL, etc. BTC, ETH
trade.xyz perps BRENTOIL, GOLD, NATGAS, SP500, NVDA, etc. xyz:BRENTOIL, xyz:GOLD

The SDK is configured to load all trade.xyz markets automatically. You can trade oil, gold, equities, and commodities alongside crypto.


Featured Strategy: Bitcoin Power Law

The flagship strategy. Based on the Bitcoin Power Law model — BTC price follows a power-law corridor over long timeframes.

  • Calculates "floor" and "ceiling" from the model
  • Near the floor: increase leverage (accumulate)
  • Near the ceiling: decrease leverage (take profit)
  • Rebalances hourly by default
hl daemon add power_law_btc -i BTC-PERP -t 3600
hl daemon start --tier rebalance

Follow via vault: HWM Opportunistic MOE on Hyperliquid


The Research Loop (How the Agent Improves)

This is not a static bot. The agent learns from every trade:

data/research/
  trades/         # Every trade with thesis, outcome, lesson learned
  market_notes/   # Market analysis snapshots
  signals/        # Signal log for future backtesting
  learnings.md    # What worked, what didn't — updated after each trade

After each trade closes, the agent reviews: Was the thesis right? Was the entry timing good? Did position sizing protect us? The lesson feeds back into how it analyzes the next trade. Over time, the codebase evolves — the agent proposes improvements, you approve, and we push to GitHub.


How It's Built (Honestly)

We frankensteined this together. Here's what came from where:

Component Origin What We Did
Strategy engine + CLI Forked from Nunchi's agent-cli Stripped all fees, telemetry, external dependencies. Kept the strategy framework.
Bitcoin Power Law Built from scratch Index-fund-style BTC rebalancer. The one strategy we trust to run fully automated.
Daemon layer Inspired by Hummingbot Tick-based iterator loop. 10 iterators. Simpler than Hummingbot — no Cython, no event bus.
Key management Open Wallet Standard + macOS Keychain AES-256-GCM encrypted vault + Keychain dual-write.
Telegram Built from scratch Two-way: alerts out, commands in. Secrets in Keychain, not config files.
Profit lock Built from scratch Auto-sweeps 25% of profits. Capital protection on autopilot.
Liquidity guard Built from scratch Time-of-day regime detection. Reduces size on weekends/after-hours.
Historical data Built from scratch SQLite candle cache, HL API fetcher, backtest engine.
MCP tools Model Context Protocol 16+ tools for AI agent integration.

What You Should Know (Transparency)

This is experimental software

We're trading real money with it. But it's new, the codebase is evolving, and we're learning as we go. Use it at your own risk. Start on testnet. Start with --tier watch. Graduate slowly.

Security is a real concern

  • API wallets protect you — the bot can never withdraw your funds
  • Keys are encrypted — OWS vault (AES-256-GCM) + macOS Keychain
  • Telegram tokens are in Keychain — not plaintext config files
  • But: if someone gets root on your machine, they can access your Keychain. If you're running on a server, consider hardware security modules
  • We never commit secretsdata/ is gitignored, pre-commit hooks scan for key patterns

The AI agent is powerful but imperfect

  • Claude Code makes autonomous trading decisions. It can and will make mistakes
  • Data manipulation is real in wartime markets — the agent can be fooled by spoofed data
  • The agent improves over time but isn't infallible
  • Always keep a mental stop on your total account exposure
  • The profit lock mechanism is your safety net — it forces partial profit-taking automatically

trade.xyz markets are different

  • BRENTOIL, GOLD, etc. are builder-deployed perps — they require isolated margin (no cross)
  • They're not in the native Hyperliquid universe — the SDK needs perp_dexs=["", "xyz"] to see them
  • Liquidity is lower than native perps — spreads are wider, stop hunts are more common
  • The contract specs (max leverage, tick size) may differ from what you expect

Commands Reference

# Daemon
hl daemon start [--tier watch|rebalance|opportunistic] [--tick 60] [--mock] [--mainnet]
hl daemon stop | status | once
hl daemon tier [watch|rebalance|opportunistic]
hl daemon strategies | add | remove | pause | resume

# Direct strategy execution
hl run <strategy> -i <instrument> [--tick N] [--mock] [--mainnet]

# Monitoring
hl status [--watch]
hl account

# Data & Backtesting
hl data fetch | stats | export
hl backtest run -s <strategy> -c <coin> -d <days>

# Key management
hl keys import | list | migrate

# Scanning
hl radar once | run
hl pulse once | run

# Infrastructure
hl setup check
hl mcp serve
hl strategies [--all] [--advanced]

Telegram Commands

Send these to your bot from your phone:

Command What It Does
/status Portfolio snapshot — positions, P&L, equity
/price Current prices for watched instruments
/help List available commands
(free text) Forwarded to Claude for analysis/execution

Development

pip install -e ".[dev]"
pytest tests/ -v

What We're Working On

  • Custom index vaults (self-rebalancing baskets — replacing index funds)
  • Deeper Hummingbot integration (their connectors are battle-tested)
  • More sophisticated entry timing using order flow analysis
  • Better backtesting with trade.xyz historical data
  • Multi-account orchestration (vault + main + sub-accounts)
  • Mobile-first experience via Telegram + Claude Code

Contributing

This is an independent project. PRs welcome. If you build something interesting, open an issue and let's talk.


Heritage

Forked from Nunchi's agent-cli. Credit to their team for the strategy engine and CLI framework. We stripped fees and telemetry, then built on top: the daemon, Bitcoin Power Law, key management, trade.xyz support, Telegram integration, profit locking, liquidity-aware risk, and AI agent tooling.


MIT License • Built by humans and Claude • Not financial advice

About

Autonomous Hyperliquid trading agent. 23 strategies orchestrated by APEX. No fees, no telemetry, runs on your own hardware.

Topics

Resources

License

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors