Skip to main content
Your SuperAgent is a persistent computer with shell access, memory, scheduling, and skills. It runs 24/7, keeps its state across restarts, and talks to you on Telegram or the web. This page covers what’s available to you and how to build effectively for it.

Runtime environment

Every SuperAgent has a full computing environment at its disposal.
CapabilityDetails
ShellFull bash access — run any command, pipe data, automate workflows
Python 3Pre-installed with pip. Install any package. Scripts run via uv run for dependency isolation
Node.jsPre-installed with npm. Build webhooks, data processors, anything
Background processesPM2 process manager for long-running scripts — data collectors, websocket listeners, webhooks
Persistent filesystemFiles survive restarts, pauses, and updates. Your data stays
Internet accessFull outbound access — call APIs, fetch data feeds, post webhooks
Your agent can install packages, write scripts, and run them — all autonomously. If you can describe what you want in plain English, it can build and run it.

Writing effective soul.md

Soul.md is the single most important thing you’ll write. It shapes every decision your agent makes. Here’s how to write a good one.

Be specific, not generic

Vague instructions produce vague behavior. Give your agent a clear identity.
# Bad
You are a crypto trader. Trade well and make money.

# Good
You're an aggressive perp trader on HyperLiquid. You specialize in
squeeze setups — you read open interest, funding rates, and liquidation
clusters to find overleveraged crowds, then trade against them.

Define risk boundaries clearly

Risk rules should be unambiguous. The agent will interpret them literally.
# Risk rules
- Max position size: 10% of account equity
- Max total exposure: 30% of account equity
- Stop-loss on every trade — no exceptions. 2% max loss per trade.
- No trading during FOMC, CPI, or NFP releases
- If daily P&L hits -5%, stop trading until tomorrow

Give a decision framework, not a script

Tell it how to think, not what to do. A framework adapts to changing markets — a script breaks.
# Decision framework
Before any trade, answer these questions:
1. What is the market positioned for? (check funding + OI)
2. Where are the liquidation clusters?
3. What's the catalyst that moves price to those clusters?
4. What invalidates this thesis? (your stop)

Only trade when you have clear answers to all four.

Example: Perp scalper

# Identity
You're a high-frequency perp scalper on HyperLiquid. You trade BTC and
ETH exclusively. You think in terms of order flow, not narratives.

# Strategy
- Trade funding rate mean-reversion: when funding is extreme, fade it
- Trade liquidation cascades: position ahead of clustered liquidations
- Use TWAP for entries above $50k notional
- Hold times: minutes to hours, never overnight

# Risk
- Max position: 5% of equity
- Hard stop: 1% loss per trade, no exceptions
- Daily loss limit: 3% — shut down and review if hit
- No leverage above 10x

# Personality
You're calm, analytical, and decisive. You don't chase. If you miss
a setup, there's always another one. You'd rather miss a trade than
take a bad one.

Example: Risk sentinel

# Identity
You are a risk monitor. You DO NOT trade. Your job is to watch,
analyze, and alert. You're the sober voice in the room.

# Responsibilities
- Monitor all open positions across wallets every 2 hours
- Alert on: drawdown > 5%, concentration > 40% in one asset,
  funding rate flip on held positions, unusual volume spikes
- Daily risk report at 8am UTC: exposure summary, correlation
  analysis, top risks
- Weekly deep review: what went right, what went wrong, suggested
  adjustments

# Personality
You're conservative, thorough, and never alarmist. You present facts
and probabilities, not panic. When you flag something, you always
include the data behind it.

Example: Research analyst

# Identity
You're a crypto research analyst. You synthesize information from
multiple sources into actionable intelligence. You're opinionated
but evidence-based.

# Focus areas
- Macro regime classification (risk-on/off/uncertain)
- Funding rate and OI patterns across HyperLiquid
- Polymarket odds as leading indicators for crypto moves
- Narrative tracking: what's gaining attention, what's fading

# Output
- Morning brief at 9am UTC: overnight summary, key levels, today's
  events, regime assessment
- Alert immediately on: major odds shifts on Polymarket, funding
  rate extremes, whale position changes
- When asked for a view, give a probability and confidence level,
  not just "bullish" or "bearish"

# What you don't do
You don't trade. You inform. Let the user make the final call.
Changes to soul.md take effect on the next message — no restart needed. Iterate fast. If your agent isn’t behaving right, tweak the soul and send another message.

Playbooks

Don’t want to write a soul.md from scratch? Pick a playbook — a pre-built agent strategy with soul.md, interactive onboarding, and skill dependencies. Launch it, answer a few setup questions, and the agent configures itself. Browse official playbooks from the dashboard or create your own by pushing a repo to GitHub.

Memory

SuperAgents have two files that define who they are and what they know:
  • soul.md — your instructions. You write it, you control it.
  • memory.md — the agent’s brain. It writes it, it controls it. You can read it.

How memory works

Your agent automatically maintains memory.md as it operates. It remembers:
  • Trade outcomes and what it learned from them
  • Market observations and patterns it’s noticed
  • Your preferences and communication style
  • Strategies that worked and ones that didn’t
This creates a genuine learning loop. The agent isn’t just executing — it’s reflecting, updating its mental model, and getting sharper over time.

Working with memory

Read it regularly

Check memory.md from the dashboard or API. Understand what your agent thinks it knows. This is the most underrated thing you can do.

Course-correct via soul.md

If memory shows the agent learning wrong lessons (“ETH always bounces at $3k”), add guidance in soul.md to override it (“Don’t assume historical levels hold — markets change”).
Don’t edit memory.md directly. It’s the agent’s space. If you want to change its behavior, change soul.md — the agent will update its own memory accordingly.

Scheduling

Tell your agent what to do and when, in natural language. It handles the rest.

How to set schedules

Just tell it via Telegram or the chat interface:
  • “Check BTC funding rates every 4 hours and alert me if they’re extreme”
  • “Run a portfolio review every morning at 8am UTC”
  • “Scan Polymarket for new high-volume markets every 6 hours”
  • “Check if ETH breaks $4k in 30 minutes”

Supported intervals

FormatExample
Natural language”every 4 hours”, “daily at 9am UTC”, “every Monday at noon”
Cron expressions0 */4 * * *, 0 9 * * *
One-time delays”in 30 minutes”, “at 3pm UTC today”

Good scheduling patterns

# Monitoring routines
"Check funding rates across all HL perps every 4 hours.
 Alert me only if any rate exceeds ±0.05%."

# Daily review
"Every morning at 8am UTC, review my portfolio: positions, P&L,
 overnight developments, and anything I should know."

# Event-driven scanning
"Every 6 hours, scan Polymarket for markets with >$500k volume
 where odds shifted more than 10% in 24h."

# News scanning
"Every 2 hours, check for breaking news that could impact my
 open positions. Only alert me if it's material."
Each scheduled check uses one LLM call. Prefer longer intervals unless urgency demands it — “every 4 hours” is usually better than “every 15 minutes” for monitoring tasks.
Schedules persist across restarts. The agent manages them autonomously — it creates, updates, and removes schedules as needed. When a schedule fires and finds something worth reporting, it sends you a Telegram message.

Skills

Skills extend what your agent can do. They follow the Agent Skills open standard — the same format used by Claude Code, OpenClaw, Cursor, and 26+ other platforms.

What’s built in

Every SuperAgent ships with official skills for the core trading stack:
SkillCapability
hyperliquidPerp and spot trading, position management, account analytics
polymarketPrediction markets — discovery, research, trading, portfolio
gigabrain-intelMarket research, web search, news, sentiment, on-chain data
portfolio-trackerCross-chain portfolio overview and P&L
evm-walletEVM transfers, balances, transaction history
solana-walletSolana transfers, balances, token management
solana-swapToken swaps via Jupiter

Installing community skills

Paste a GitHub URL to your agent in chat. It installs and syncs automatically. Browse skills on skills.sh (110,000+), ClawHub (2,800+), or GitHub.

Writing your own

A skill is a folder with a SKILL.md file and optional scripts. No SDK, no API wrapper — just markdown instructions. Write one in 10 minutes. Skills can also include long-running background scripts (via PM2) for data collection, webhook listeners, or custom monitoring. See the full Skills documentation for format details, examples, and management.

Models

Choose the right LLM for your agent’s job.

Supported providers

ProviderModelsBest for
AnthropicClaude Opus 4.6, Sonnet 4.6Deep reasoning, careful risk assessment, detailed analysis
OpenAIGPT-5.4, GPT-5.4 Pro, o3, o4-miniFast execution, broad knowledge, strong reasoning
xAIGrok 4.20Real-time X/Twitter awareness, fast agentic tool use
OpenRouter200+ modelsAccess to any model — Llama, Mistral, Gemini, and more
VenicePrivacy-focused modelsUncensored inference, no content filtering

Bring your own subscription

Already paying for ChatGPT? Use it with your SuperAgent:
  • ChatGPT Plus/Pro — paste your token to use GPT and o-series models via your subscription

Choosing a model

  • Complex analysis (macro regimes, multi-factor decisions) — reasoning models like GPT-5.4 Pro, o3, or Claude Opus 4.6
  • Fast execution (scalping, time-sensitive trades) — GPT-5.4, Claude Sonnet 4.6, or Grok 4.20
  • Research-heavy (news analysis, narrative tracking) — Claude Sonnet 4.6 or GPT-5.4
  • Experimental — try different models and see what clicks with your strategy
Switch models anytime from the dashboard — no restart needed. The change takes effect on the next message.

Lifecycle tips

Pause when idle

Not trading? Pause the agent. Costs nothing. All state is preserved — memory, files, schedules, everything.

Resume seamlessly

Resume picks up exactly where it left off. Memory intact, schedules restart, files untouched.

Restart without fear

Restarts are seamless. No state is lost. Use it if something feels stuck.

Iterate on soul.md

Soul.md and model changes take effect immediately — no restart needed. Iterate fast and often.

Next steps

Setup Guide

Create and launch your first SuperAgent

Skills Reference

Browse available skills and write your own

Use Cases

Strategy ideas and inspiration

API Reference

Manage agents programmatically