Synthesis Hackathon submission. Built on IronClaw fork. 26 tools: Locus payments (6), Uniswap DeFi (3), onchain reads (3), security scanning (2), wallet analytics (3), OpenRouter LLM (2), self-hosted scraping (3), portfolio (2), synthesis (2). 3 defense hooks: spending limits ($5/tx, $25/session), contract guard (SCONE-bench vulnerability patterns), credential guard (blocks private key/seed phrase leaks). Payment system (Polar.sh), landing page, license key gating (Free/Pro/ Unlimited tiers), brand rename from IronClaw to Steelbot. 9,669 lines of code. 3,665 tests passing. Zero clippy warnings. Real USDC transactions on Base chain. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Steelbot
The AI agent that pays its own way.
What is Steelbot?
Steelbot is a secure, self-expanding AI assistant with built-in cryptocurrency payments. It manages USDC wallets on Base (Ethereum L2), executes DeFi trades via Uniswap, scans smart contracts for vulnerabilities using SCONE-bench patterns, and defends against prompt injection attacks -- all while enforcing strict per-transaction and per-session spending limits. Built in Rust for memory safety and performance, Steelbot is designed so that AI agents can handle real money without compromising security.
Features
- 21 built-in tools -- payments, DeFi swaps, on-chain reads, portfolio analytics, security scanning
- 3 defense hooks -- spending limits, contract guard, credential leak prevention
- BYOW (Bring Your Own Wallet) -- connect your Locus wallet for USDC on Base
- Multi-channel access -- CLI (full TUI), Web UI, HTTP webhooks
- Tier-based licensing -- Free / Pro ($19/mo) / Unlimited ($49/mo)
- Self-expanding tools -- build new WASM tools on the fly without restarting
- Persistent memory -- hybrid full-text + vector search across sessions
- Built with Rust -- single binary, async tokio runtime, WASM sandbox isolation
Quick Start
Install (shell script)
curl -sSL https://steelbot.dev/install.sh | bash
Install (cargo)
git clone https://github.com/ArhansSubasi/ironclaw.git
cd ironclaw
cargo build --release
# Binary at target/release/ironclaw
Install (Docker)
docker pull steelbot/steelbot:latest
docker run --env-file .env -p 3000:3000 steelbot/steelbot:latest
Configure
Create a .env file (or copy from .env.example):
cp .env.example .env
Required for payment tools:
LOCUS_API_KEY=claw_... # Enables USDC payment tools
STEELBOT_WALLET_ADDRESS=0x... # Your wallet address on Base
BASE_RPC_URL=https://mainnet.base.org # Base chain JSON-RPC
LLM provider (pick one):
# NEAR AI (default)
LLM_BACKEND=nearai
# Anthropic
LLM_BACKEND=anthropic
ANTHROPIC_API_KEY=sk-ant-...
# OpenAI / OpenRouter / any OpenAI-compatible
LLM_BACKEND=openai_compatible
LLM_BASE_URL=https://openrouter.ai/api/v1
LLM_API_KEY=sk-or-...
LLM_MODEL=anthropic/claude-sonnet-4
# Local (Ollama)
LLM_BACKEND=ollama
OLLAMA_MODEL=llama3.2
Run
# First-time setup wizard
steelbot onboard
# Start the agent
steelbot run
# With debug logging
RUST_LOG=ironclaw=debug steelbot run
Tools
Payments (Locus API)
| Tool | Description | Approval |
|---|---|---|
locus_balance |
Check USDC balance on Base | Never |
locus_send |
Send USDC to any address | Always |
locus_transactions |
View transaction history | Never |
locus_checkout |
Pay merchant checkout sessions | Always |
locus_wrapped |
Pay-per-use API calls (Firecrawl, Gemini, Exa) | Auto |
locus_x402 |
Call x402 pay-per-call endpoints | Auto |
DeFi (Uniswap + DexScreener)
| Tool | Description | Approval |
|---|---|---|
uniswap_quote |
Get swap quotes for any token pair on Base | Never |
token_search |
Search tokens by name/symbol via DexScreener | Never |
token_price |
Real-time token price and pool data | Never |
On-chain Reads (Base RPC)
| Tool | Description | Approval |
|---|---|---|
erc20_balance |
Check any ERC-20 token balance | Never |
erc20_info |
Token metadata (symbol, decimals, total supply) | Never |
base_rpc |
Raw JSON-RPC calls (read-only allowlist) | Never |
Wallet Analytics
| Tool | Description | Approval |
|---|---|---|
claw_wallet |
Multi-token holdings analysis | Never |
claw_gas |
Real-time gas cost estimation | Never |
claw_allowances |
ERC-20 approval audit (security) | Never |
Portfolio & Marketplace
| Tool | Description | Approval |
|---|---|---|
claw_portfolio |
Aggregate portfolio view (balance + holdings + history) | Never |
claw_services |
Service marketplace (list, price, fulfill, revenue) | Fulfill: Auto |
Defensive Security
| Tool | Description | Approval |
|---|---|---|
claw_scan |
Contract bytecode vulnerability scanner (7 SCONE-bench patterns) | Never |
claw_risk |
Pre-transaction risk assessment | Never |
Hackathon
| Tool | Description | Approval |
|---|---|---|
synthesis_team |
Hackathon team management | Mutating: Auto |
synthesis_status |
Registration and verification status | Never |
Defense Hooks
| Hook | Trigger Point | What It Does |
|---|---|---|
LocusSpendingHook |
BeforeToolCall | Enforces $5/transaction limit, $25/session cap, emergency kill switch |
ContractGuardHook |
BeforeToolCall | Warns on interactions with unverified contracts, checks trusted allowlist |
CredentialGuardHook |
BeforeOutbound + BeforeToolCall | Blocks private keys, seed phrases, and API keys from leaking in any output |
Security
Steelbot was built in direct response to two Anthropic research findings:
SCONE-bench (Dec 2025): AI exploited 51% of 405 vulnerable smart contracts, simulating $550M in stolen funds. Found 2 novel zero-days in 2,849 fresh contracts.
Anthropic Threat Report (Aug 2025): Claude Code was weaponized for extortion targeting 17 organizations. AI-powered ransomware sold for $400--$1,200.
Steelbot's defense stack:
- Contract bytecode scanner checks 7 SCONE-bench vulnerability patterns before any interaction
- Spending hook caps financial exposure at $5/transaction and $25/session
- Credential guard prevents exfiltration of private keys, seed phrases, and API keys
- All payment operations require explicit human approval
- Read-only RPC --
eth_sendRawTransactionis blocked - Trusted contract allowlist (USDC, WETH, Uniswap, Aerodrome, Safe, Multicall3)
- WASM sandbox isolates untrusted tool execution with capability-based permissions
- Prompt injection detection with pattern matching and content sanitization
Pricing
| Free | Pro | Unlimited | |
|---|---|---|---|
| Price | $0/mo | $19/mo | $49/mo |
| Tools | 10 | 21 | 21 |
| Hooks | 1 (credential guard) | 3 | 3 |
| Tx limit | $1/tx | $5/tx | $50/tx |
| Session cap | $5/session | $25/session | $500/session |
| Channels | CLI only | CLI + Web | CLI + Web + Webhooks |
| Memory | 100 docs | 10,000 docs | Unlimited |
| Custom tools | -- | 5 WASM tools | Unlimited |
| Support | Community | Priority |
Architecture
Steelbot Agent
+-------------------------------------------------+
| |
| Channels: CLI (TUI) | Web UI | HTTP Webhooks |
| | |
| Agent Loop (LLM + 21 Tools) |
| | |
| +--------+--------+--------+--------+------+ |
| | Locus | Uniswap| OnChain| Wallet | Sec | |
| | Pay(6) | Swap(3)| Read(3)| Ana(3) | (2) | |
| +---+----+---+----+---+----+---+----+--+---+ |
| | | | | | |
| +---------------------------------------------+
| | 3 Defense Hooks |
| | Spending | Contract | Credential |
| | Policy | Guard | Guard |
| +---------------------------------------------+
| | |
| Base Chain (USDC) |
+-------------------------------------------------+
Development
# Format
cargo fmt
# Lint (zero warnings required)
cargo clippy --all --benches --tests --examples --all-features
# Run tests
cargo test --lib
# Run with debug logging
RUST_LOG=ironclaw=debug cargo run
# Build Docker image
docker build -t steelbot:latest .
Documentation
- LLM Providers Guide -- configure Anthropic, OpenAI, Ollama, and 10+ other backends
- Telegram Setup -- channel configuration and DM pairing
.env.example-- all environment variables with documentation
License
Licensed under either of:
- Apache License, Version 2.0 (LICENSE-APACHE)
- MIT License (LICENSE-MIT)
at your option.
Contributing
Contributions are welcome. Please follow these guidelines:
- Fork and branch -- create a feature branch from
staging - Code style -- run
cargo fmtand ensure zerocargo clippywarnings - Tests -- add tests for new functionality; every bug fix needs a regression test
- No panics -- no
.unwrap()or.expect()in production code (tests are fine) - PR to staging -- all PRs target the
stagingbranch; promotion tomainis automated
See CLAUDE.md for detailed development guidelines.
