Skip to main content
Glama

Aegis

自主 DeFi 代理的安全层。 | 网站 | 文档

在链上进行交易的 AI 代理无法区分合法代币和蜜罐。Aegis 解决了这个问题。它是一个任何代理都可以接入的 MCP 服务器,并由强制执行安全检查的链上合约提供支持。

在代理进行兑换之前,Aegis 会扫描目标合约、模拟交易,并返回一个简单的“通过/不通过”结果。如果合约有 99% 的卖出税或隐藏的暂停功能,代理将永远不会触碰它。

为什么存在这个项目

我们曾目睹一个代理在不到 30 秒的时间内因蜜罐代币损失了整个钱包的资金。该代币表面上看起来没问题——合约已验证、流动性充足、交易活跃。但代码中隐藏着 99% 的卖出税,以及一个在虚假的 renounceOwnership() 函数背后的隐藏所有者。

没有任何代理框架能够捕捉到这一点。所以我们构建了一个。

Related MCP server: agentshield-mcp

工作原理

Agent -> Aegis (scan + simulate + decide) -> Chain
  1. 代理通过 MCP 连接到 Aegis(一行配置)

  2. 在任何兑换/授权/转账之前,代理调用 assess_risk

  3. Aegis 扫描合约源代码,模拟交易,检查蜜罐模式

  4. 返回 ALLOW(允许)、WARN(警告)或 BLOCK(拦截),并附带风险评分(0-100)

  5. 链上:AegisGateway 合约在转发交易前强制执行证明

快速开始

# Add to Claude Code
claude mcp add aegis npx aegis-defi

# Or clone and try the demo
git clone https://github.com/StanleytheGoat/aegis
cd aegis && npm install
npx tsx demo/catch-honeypot.ts

该演示部署了一个蓄意恶意的代币(99% 卖出税、虚假放弃所有权、隐藏管理员),并观察 Aegis 如何捕捉到每一个危险信号:

Aegis Risk Assessment
  Risk Score: 100/100
  Findings:
    [CRITICAL] Fake Ownership Renounce
    [CRITICAL] Asymmetric Buy/Sell Tax (99% sell)
    [CRITICAL] Sell Pause Mechanism
    [HIGH]     Hidden Max Sell Amount
    [HIGH]     Hidden Admin Functions
  Decision: BLOCK

工具

MCP 服务器 (TypeScript) - 为任何兼容 MCP 的代理提供 6 种工具:

工具

用途

scan_contract

针对 165 种已知漏洞类型的模式匹配

simulate_transaction

在分叉链上进行试运行

check_token

反蜜罐检查(可卖性、持仓集中度)

assess_risk

包含签名证明的一站式风险评估

trace_transaction

追踪每一个内部调用,扫描每个合约

search_solodit

交叉引用 5 万多条真实审计发现

智能合约 (Solidity) - 部署在 Base 主网上:

合约

地址

用途

AegisGateway

0x62c6...0fa3

任何 DeFi 交互的安全包装器。验证证明,检查风险评分。

AegisSafetyHook

0xaEE5...40C0

Uniswap v4 beforeSwap 钩子。拦截没有有效安全证明的兑换。

文档

安全性

遵循以太坊安全最佳实践构建(参考 ethskills):

  • 签名:所有签名消息中包含链 ID + 合约地址(无跨链重放)。EIP-2 s-值延展性检查。针对 address(0) 验证 ecrecover。

  • 费用数学:先乘后除。显式溢出防护。基点(而非百分比)。

  • 访问控制:Gateway 上的 OZ Ownable + ReentrancyGuard。Hook 上的不可变所有者。不可变的费用接收者。

  • 部署:Safe Singleton Factory CREATE2 部署器。源代码在 Basescan 上验证。所有权转移至 Safe 多签钱包。

  • 测试:165 个测试(42 个合约测试 + 123 个 TypeScript 测试)。针对真实 Base 主网状态的分叉测试。

测试

npm test              # TypeScript unit tests (123)
npm run test:contracts # Solidity contract tests (42)
npm run demo          # Honeypot detection demo

更新日志

v0.5.0 (当前版本)

  • Hook 证明支持 - assess_risk 现在为 Uniswap v4 受保护池返回网关和 Hook 证明

  • EVM 地址验证 - 所有 MCP 工具输入均验证正确的地址格式

  • 扩展知名合约 - Paraswap, Balancer Vault, CoW Protocol, Permit2, Uniswap V4 PoolManager

  • SDK 导出 - 证明器和 solodit 模块现可用于编程使用

  • 强化获取 - response.ok 检查,所有外部请求均有 10 秒超时

  • 安全标头和落地页 SEO 文件

v0.4.0

  • Solodit 集成 - search_solodit 工具查询来自 Cyfrin, Sherlock, Code4rena, Trail of Bits 等机构的 5 万多条真实审计发现

  • 自动丰富 - 当设置 SOLODIT_API_KEY 时,assess_risk 会将检测到的模式与真实审计发现进行交叉引用

  • 选择性 API 密钥模型 - 每个代理提供自己的 Solodit 密钥,无共享速率限制

v0.3.0

  • 165 种漏洞模式,涵盖 25 个类别(之前为 22 个)

  • 追踪级分析 - trace_transaction 工具跟踪每一个内部调用并扫描每个合约

v0.2.0

  • 22 种漏洞模式(之前为 12 个) - 变形合约、预言机操纵、MEV 三明治攻击

  • 代理技能 - 可安装的 Claude Code 技能文件

  • Flaunch SDK 集成 - Uniswap v4 池上模因币发布的安全扫描

许可证

MIT

Available Tools

4 tools
assess_riskA

Comprehensive risk assessment combining contract scanning, transaction simulation, and token checks. This is the recommended all-in-one safety check before any DeFi interaction. Returns a go/no-go recommendation.

Inputs: action (required), targetContract (required), from (required), chainId, +3 more.

check_tokenA

Check if a token is safe to trade. Detects honeypot mechanics (can't sell), concentrated holdings, fake ownership renouncement, and other scam indicators. Use this before swapping into any unfamiliar token.

Inputs: tokenAddress (required), chainId, holderAddress.

scan_contractA

Analyze a smart contract's source code or bytecode for known exploit patterns, honeypot mechanics, rug-pull signals, and security vulnerabilities. Returns a risk score (0-100) and detailed findings. Use this BEFORE interacting with any unfamiliar contract.

Inputs: source, bytecode, contractAddress, chainId.

simulate_transactionA

Simulate a transaction on a forked chain WITHOUT actually executing it. Detects reverts, abnormal gas usage, and other red flags. Use this to preview what will happen before sending a real transaction.

Inputs: from (required), to (required), data (required), chainId, +1 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.4/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose: assess_risk is a comprehensive all-in-one safety check, check_token focuses on token-specific scams, scan_contract analyzes contract code/bytecode, and simulate_transaction previews transaction outcomes. There is no overlap or ambiguity between these tools.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun pattern (assess_risk, check_token, scan_contract, simulate_transaction), using snake_case throughout. The naming is predictable and readable across the entire set.

    Tool Count5/5

    With 4 tools, this server is well-scoped for DeFi security. Each tool earns its place by covering distinct aspects of safety assessment: holistic risk, token checks, contract analysis, and transaction simulation. This count is appropriate and avoids bloat.

    Completeness5/5

    The tool set provides complete coverage for DeFi security workflows: it includes comprehensive risk assessment (assess_risk), targeted checks for tokens and contracts, and transaction simulation. There are no obvious gaps—agents can perform end-to-end safety evaluations before any DeFi interaction.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      B
      quality
      D
      maintenance
      Real-time smart contract security for autonomous AI agents, offering tools for contract verification, wallet monitoring, drain detection, threat reporting, and leaderboards.
      19
      25 npm
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Security layer for AI agents that evaluates transaction intents and returns verdicts (ALLOW/WARN/DENY) using deterministic rules, on-chain checks, and simulation.
      9 npm
      MIT