Multi-chain crypto + Web3 agent toolkit — execute on-chain operations through natural language across Ethereum, BSC, Base, Arbitrum, Optimism, Polygon, Avalanche, Linea, Scroll, Solana, Sui, Aptos, and TON.

Built as a skill pack for AI coding agents — works with Claude Code, Cursor, OpenClaw, and any agent runtime that supports skill-based tool routing. Designed to be paired with Xiaomi MiMo, Anthropic Claude, OpenAI, and other reasoning models that drive the agent loop.

A user types "swap 100 USDC ke ETH di Base" or "bridge 0.5 ETH dari Arbitrum ke Optimism via Across" — the agent identifies the task, loads the correct reference module, simulates the transaction, asks for confirmation, and broadcasts.

Every operation goes through a strict pipeline:

Intent  →  Validate (chain, contract, ownership)
        →  Compose tx (no broadcast)
        →  Simulate (eth_call / dry-run)
        →  User confirms
        →  Broadcast → return tx hash + explorer link
        →  Monitor finality

Hardcoded crypto bots break the moment a chain forks or a DEX updates its router. This pack is declarative reference + production scripts — the agent reads the relevant file at task time, composes the call from current ABI/router data, and executes. New chain support means adding one reference file, not rewriting a binary.

This pattern works well with the new generation of long-context, tool-use-capable models like Xiaomi MiMo V2.5, where the agent can hold the full reference module in context while reasoning about a multi-step swap or bridge plan.

1. User-funds-only rule. The agent only operates on wallets the user owns and confirms. It refuses to act on third-party seed phrases or credentials presented as "target wallets" or "found in a group."
2. No drainer, no scam-sybil. No drainer code, no phishing payloads, no malicious approval generators.
3. Confirm before signing. Every fund-moving transaction shows a plan first: chain, action, amount, recipient, est. gas, slippage.
4. Simulate before broadcast. Dry-run via eth_call (or chain equivalent). Honeypots, zero-output traps, and extreme slippage are caught before a single wei moves.
5. Secret hygiene. Private keys and seed phrases are never logged, printed raw, or sent to third-party services. Local encrypted storage only.

Full rules: SKILL.md and references/security.md.

hermes-crypto-agent/
├── SKILL.md                      ← Skill manifest (the agent loads this first)
├── README.md
├── references/                   ← Per-capability operating docs
│   ├── wallets.md
│   ├── swap.md
│   ├── nft.md
│   ├── sniping.md
│   ├── airdrop_automation.md
│   ├── bridge.md
│   ├── defi.md
│   ├── web3_connect.md
│   ├── monitoring.md
│   └── security.md
└── scripts/                      ← Standalone production templates
    ├── wallet_manager.py
    ├── swap_engine.py
    ├── nft_engine.py
    ├── bridge_engine.py
    ├── airdrop_runner.py
    ├── web3_connect.py
    ├── monitoring.py
    └── monitoring_advanced.py

• Claude Code — drop into your skills directory, agent auto-discovers.
• Cursor — reference from .cursorrules or paste capability tables into context.
• OpenClaw — register as a skill module in your agent config.
• Xiaomi MiMo API — long-context reasoning over the full reference set; ideal for multi-step planning.

MIT.

This is a defensive-by-default crypto operations toolkit. Do not use it to operate on wallets you do not own.