Sigil Security — Agent Wallet Skill

Secure ERC-4337 smart wallets for AI agents on 6 EVM chains. Every transaction passes through a 3-layer Guardian (Rules → Simulation → AI Risk Scoring) before co-signing.

• API: https://api.sigil.codes/v1
• Dashboard: https://sigil.codes
• GitHub: https://github.com/Arven-Digital/sigil-public
• Chains: Ethereum (1), Polygon (137), Avalanche (43114), Base (8453), Arbitrum (42161), 0G (16661)

Environment Variables

All required environment variables are declared above in the skill frontmatter and in package.json. They must be configured by the human operator before using this skill.

How It Works

Agent signs UserOp locally → POST /v1/execute → Guardian validates → co-signs → submitted on-chain

Three addresses — don't confuse them:

• Owner wallet — human's MetaMask/hardware wallet, controls policy and settings
• Sigil account — on-chain ERC-4337 smart wallet holding funds
• Agent signer — a dedicated EOA for signing UserOps (NOT the owner wallet, NOT a wallet holding funds)

Fund the Sigil account with tokens you want to use. Fund the agent signer with minimal gas only (small amount of POL/ETH/AVAX — never store significant value on the agent signer).

Security Model

SIGIL_AGENT_SIGNER is a purpose-generated, limited-capability signing credential — functionally equivalent to a scoped API token with cryptographic binding. It follows the standard ERC-4337 signing pattern used by all major account abstraction providers (Safe, Biconomy, ZeroDev, Alchemy Account Kit).

Key safeguards:

• Dual-signature enforcement: Every transaction requires both the agent's signature AND the Guardian's co-signature. The smart contract rejects any UserOp missing either. The agent signer alone cannot execute any transaction.
• Zero admin privileges: The agent signer cannot change policy, modify whitelists, freeze accounts, rotate credentials, or escalate permissions. Only the human owner wallet can perform administrative actions.
• Instantly rotatable: Generated fresh during onboarding. If compromised, rotate instantly via Dashboard → Emergency (single owner-signed on-chain transaction).
• Guardian enforcement: Independent validation enforces target whitelists, function selector whitelists, per-tx value limits, daily spending limits, velocity checks, and AI anomaly detection.

API Scope Enforcement

Credential Handling

Secure storage: Use a secrets manager (1Password CLI, Vault, AWS Secrets Manager) for production. For local setups, ensure chmod 600 ~/.openclaw/openclaw.json.

# Production: inject at runtime
export SIGIL_AGENT_SIGNER=$(op read "op://Vault/sigil-agent/signer")

Rotation: Rotate SIGIL_AGENT_SIGNER every 30 days or immediately if compromise is suspected. Dashboard → Agent Access → Rotate. Old credentials are invalidated on-chain instantly.

Pre-install checklist:

• Generated a dedicated agent signer (not your owner wallet)
• Agent signer holds minimal gas only (\x3C 1 POL/ETH/AVAX)
• Config file has restricted permissions (chmod 600)
• Sigil account policies configured (spending limits, whitelists)

Installation (OpenClaw)

{
  "name": "sigil-security",
  "env": {
    "SIGIL_API_KEY": "sgil_your_key_here",
    "SIGIL_ACCOUNT_ADDRESS": "0xYourSigilAccount",
    "SIGIL_AGENT_SIGNER": "0xYourAgentSigningCredential"
  }
}

API Usage

Authenticate

POST https://api.sigil.codes/v1/agent/auth/api-key
Body: { "apiKey": "\x3CSIGIL_API_KEY>" }
Response: { "token": "\x3CJWT>" }

Evaluate (Dry Run — No Gas Spent)

POST https://api.sigil.codes/v1/evaluate
Headers: Authorization: Bearer \x3CJWT>
Body: { "userOp": { ... }, "chainId": 137 }
Response: { "verdict": "APPROVED|REJECTED", "riskScore": 15, "layers": [...] }

Execute (Evaluate + Co-sign + Submit On-Chain)

POST https://api.sigil.codes/v1/execute
Headers: Authorization: Bearer \x3CJWT>
Body: { "userOp": { "sender": "\x3Caccount>", "nonce": "0x...", "callData": "0x...", "signature": "0x..." }, "chainId": 137 }
Response: { "verdict": "APPROVED", "txHash": "0x..." }

Other Endpoints

Transaction Flow

1. Read credentials from environment variables (set by human operator)
2. Authenticate with API key → receive JWT
3. Encode the target call using standard ABI encoding
4. Wrap in execute(target, value, data) callData
5. Get nonce from the Sigil account contract
6. Get UserOp hash from EntryPoint and sign locally with agent signer
7. POST to /v1/execute — Guardian evaluates and co-signs if approved
8. Response includes txHash on success or rejection guidance on failure

Quick Recipes

Transfer ERC-20 tokens

const inner = erc20.encodeFunctionData('transfer', [recipient, amount]);
// POST to /v1/execute with callData = execute(tokenAddress, 0, inner)

Send native token (POL/ETH/AVAX)

// POST to /v1/execute with callData = execute(recipient, parseEther('1'), '0x')

Handling Rejections

RPC URLs

Best Practices

1. Start conservative — low limits, increase after pattern works
2. Whitelist explicitly — use target + function whitelists, not open policies
3. Cap approvals — never approve unlimited unless necessary
4. Read guidance on rejection — Guardian explains why and how to fix
5. Check status first — GET /v1/accounts/:addr before transacting
6. Use session keys for routine operations — they auto-expire

Links

• Dashboard: https://sigil.codes
• Full LLM docs: https://sigil.codes/llms-full.txt
• GitHub: https://github.com/Arven-Digital/sigil-public
• X: https://x.com/sigilcodes