mcp-growi
Enables AI agents and autonomous bots to interact programmatically with the GrowiHFVault smart contract on Arbitrum.
README
mcp-growi
MCP (Model Context Protocol) server for Growi Finance — enables AI agents and autonomous bots to interact programmatically with the GrowiHFVault smart contract on Arbitrum.
Security
Read this before configuring a keystore or private key.
mcp-growi exposes execute_transaction, which signs and broadcasts Arbitrum transactions with the locally configured wallet. Inside any MCP client (Claude Desktop, Claude Code, etc.) the LLM has full authority over what gets signed — every tool call is initiated by the model in response to your prompts and anything else it has access to.
This makes the MCP a "confused deputy": if untrusted content reaches the model — a web page fetched by another tool, a document opened in the session, a sibling MCP server controlled by an attacker — a prompt-injection payload inside that content can instruct the model to drain the configured wallet.
Recommendations
- Do not configure
KEYSTORE_PATHorPRIVATE_KEYin a session that has access to untrusted content. If the same client has browsing tools, document-readers, or any other MCP server, treat the session as compromised by default. - Use a dedicated "bot" wallet. Create a fresh Ethereum address solely for
mcp-growiand fund it only with the amount you intend to deposit. Never reuse your main custody wallet. - Keep ETH for gas low. A few dollars' worth is enough — there is no upside to holding more on the bot wallet than the next operation needs.
- Run
mcp-growi@latest. When a security release is published, the built-in version check blocks all operations until you upgrade. Pinning to an older version defeats this.
What this server does enforce
The server has its own defenses against the most direct drain paths, but they cannot prevent every form of misuse:
- Allowlist.
execute_transactiononly signs calls to four(contract, selector)pairs —USDC.approve,GWHF.approve,vault.deposit,vault.withdraw. A prompt-injection payload callingUSDC.transfer(attacker, balance)is rejected before signing. - Decoded validation.
approvecalls are accepted only when the spender is the GrowiHFVault address.depositandwithdrawamounts must be greater than zero. - Version gate. All read and write tools refuse to operate when the local MCP version cannot be verified against the npm registry, or does not match the latest published version.
These guards close the easy attacks but do not make the keystore safe in arbitrary sessions. Treat it as a hot wallet, not custody.
Contract details
| Field | Value |
|---|---|
| Network | Arbitrum One |
| Contract | 0x928ed672e6eabb7a565c5eb9aac15e3cf6a18388 |
| Vault token | GWHF |
| Deposit token | USDC |
The contract is verified on Arbiscan — you can check the source code and ABI at: https://arbiscan.io/address/0x928ed672e6eabb7a565c5eb9aac15e3cf6a18388#code
The ABI is included in this package (src/contract/abi.ts) so no external calls are needed at runtime.
Project structure
src/
├── index.ts # Server entrypoint + npm version check
├── contract/ # On-chain interaction layer (viem)
│ ├── abi.ts # GrowiHFVault + ERC20 ABIs
│ ├── client.ts # Arbitrum public client + contract/token/HL addresses
│ └── index.ts
├── locks/
│ └── lock-period.ts # Arbitrum deposit lock + Hyperliquid 24h withdraw lockup
└── tools/ # MCP tool definitions
├── read.ts # Read-only on-chain queries
├── write.ts # Transaction prep + signing (keystore/privkey)
└── index.ts
Setup — Claude Desktop / Claude Code
Add the following to your Claude MCP config (claude_desktop_config.json or .mcp.json):
{
"mcpServers": {
"growi": {
"command": "npx",
"args": ["-y", "mcp-growi@latest"],
"env": {
"KEYSTORE_PATH": "/path/to/your/wallet.json",
"KEYSTORE_PASSPHRASE": "your_passphrase"
}
}
}
}
Important: use
mcp-growi@latest(not justmcp-growi) to ensure the MCP auto-updates every time Claude starts. If your version is outdated, or the npm registry is unreachable, all tools (read and write) will be blocked until verification succeeds.
Alternative — plain private key (less secure):
"env": { "PRIVATE_KEY": "0x..." }
Optional — custom RPC endpoint:
"env": { "RPC_URL": "https://arb-mainnet.g.alchemy.com/v2/YOUR_KEY" }
Restart Claude Desktop/Code after saving. The tools will be available once the version check passes — if calls return a version-block error, see the Security section above.
Setup — development (from source)
npm install
npm run build
npm start
Tech stack
- TypeScript + Node.js
- @modelcontextprotocol/sdk — MCP server implementation
- viem — Ethereum/Arbitrum RPC interactions
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。