Design and enforce strict transaction policies, execution guardrails, simulations, and human approval gates for autonomous agents.
Allow AI agents to reason about on-chain actions while completely isolating signing authority behind strict external policy engines. Enforces fail-closed rules and defenses against prompt injection.
JSON policy configuration establishing chain, contract, selector, and value limits
Independent policy evaluation gate decoupled from LLM context
Preflight simulation requirements and state validation checks
Negative security test suite verifying denied transactions fail closed
Separates read-only information tools from state-changing transaction tools, identifying the exact signer boundary.
Defines allowed chain IDs, target contract addresses, 4-byte selectors, maximum native/token values, and approval tiers.
Ensures the validation gate executes outside the LLM reasoning runtime so adversarial inputs cannot bypass it.
Wires preflight simulation (via eth_call or Tenderly/Alchemy) and human approval hooks for actions exceeding policy limits.
Executes test suite with unapproved selectors, excessive values, and injected prompts to prove the system fails closed.
Once the skill is installed, trigger it in Claude Code, Cursor, or your AI assistant with:
“Use the web3-agent-tx-safety skill to audit our Telegram bot's transaction generation pipeline. We need a strict JSON policy that only permits calling transfer() on USDC on Arbitrum up to $100 per transaction.”Join the developer mailing list for new version releases, test scripts, and new Web3 skills. No spam.