A team of 10 specialized AI agents for Ethereum smart contract development.
Built for Claude Code. Zero configuration. Natural language interface.
You write in plain English. The orchestrator routes your intent to the right agent pipeline. Agents coordinate automatically — no commands to remember, no toolchain configuration.
No flags, no commands. "Audit this contract", "build a staking contract with ERC4626", "deploy to Arbitrum Sepolia" — eth-agents understands intent.
A coordinator agent identifies the skill, selects the right agents in the right order, and manages handoffs automatically. You see progress in the Claude Code status bar.
Each agent has a specific role, defined tools, and output constraints. They produce structured artifacts — audit reports, test files, gas diffs, deployment receipts.
Mainnet deployments always halt for your explicit confirmation. All agents report what they did and why. You stay in control.
Each agent is a markdown file with a defined role, allowed tools, output constraints, and handoff rules. Model routing keeps costs low — Haiku for simple tasks, Sonnet for implementation, Opus for security.
Skills are entry points. They activate coordinated agent pipelines. You trigger them by describing intent — no syntax to memorize.
eth-agents generates structured, actionable audit reports. Every finding includes severity, impact, location, and a concrete remediation with code.
Privileged function can redirect funds to any arbitrary address. Direct fund loss vector.
Cross-function reentrancy possible via callback-capable token. Risk of double refund.
Raw transfer calls will revert with non-standard ERC-20 tokens, bricking deposit/refund flows.
Lost key locks all rooms. Compromised key gives full prize control. No Ownable2Step.
No enforcement that both players stake the same amount. Griefing via front-running possible.
Validator manipulation (~12–15s) represents ~33% of the window. Increase timeout to mitigate.
Standard off-chain indexing tools won't detect initial owner assignment.
eth-agents is engineered to minimize token usage without sacrificing quality. Three complementary mechanisms work in layers.
| Layer | Mechanism | Savings |
|---|---|---|
| Model routing | Haiku for simple tasks, Sonnet for implementation, Opus only for security and full-protocol pipelines | ~35% |
| Prompt caching | Static agent role definitions are cached after first invocation (each agent file exceeds Anthropic's 1024-token auto-cache threshold) | ~15% |
| Output constraints | Per-agent token limits, artifacts always before prose, structured outputs reduce verbose explanations | ~12% |
Install the plugin, navigate to any Ethereum project, and start describing what you want. eth-agents auto-detects Foundry or Hardhat, your frontend framework, and your Web3 library.