Agentpay Sentinel Mcp

io.github.Rumblingb由 GitHub 帳號 Rumblingb 發布v1.0.0更新於 Oct 24, 2026

Security guardrails for AI agent payments

發布方自己的服務：AgentPay — Pay-per-call AI Microservices

已驗證STDIO僅桌面金融安全與監控

服務詳細資訊

識別碼: io.github.Rumblingb/agentpay-sentinel-mcp
傳輸: STDIO
執行環境: Desktop only
信任: Verified
發布於: Sep 16, 2026

安全提示

MCP 服務會接收對話中的工具呼叫請求，並可能執行外部操作。啟用前請先了解該服務能做什麼，並只授予它所需的權限。

註冊中心已驗證此條目所屬的 GitHub 帳號 Rumblingb。這並不表明該帳號就是產品方。

SourceWeft 的檢查只涵蓋條目是否完整、端點是否回應；不審查伺服器的程式碼，也不審查其工具的行為。

在 README 中新增徽章

向讀者展示該 MCP 服務已收錄於 SourceWeft，徽章連結到本頁。

[![Listed on SourceWeft](https://sourceweft.com/badge/mcp/io.github.Rumblingb%2Fagentpay-sentinel-mcp.svg)](https://sourceweft.com/mcp/io.github.Rumblingb%2Fagentpay-sentinel-mcp?utm_source=readme&utm_medium=badge)

概覽

AI 產生的概覽

一個本機看門狗 MCP，對 AI 代理的付款請求執行預檢安全檢查，回傳 PASS 或 BLOCKED 結論。

可以做什麼: 對付款交易執行九項檢查（權杖雜湊、預算上限、商家白名單、受限類別、到期、速率限制、金額不符、重放、撤銷），回傳 PASS/BLOCKED 結論與稽核雜湊
撤銷權杖，讓後續稽核一律攔截；為合法重試從重放儲存中清除 nonce
驗證一串稽核雜湊是否構成完整鏈條
模擬權杖偽造、預算溢出、重放、商家冒充等攻擊向量，回傳嚴重程度、防禦與偵測方式
適用情境: 當代理被允許發起付款，而你希望在每次付款前用本機建議性關卡檢查政策、預算與權杖完整性時適用。也適合用來了解常見付款攻擊向量會如何被偵測。
執行需求: 以本機 Python 程序執行：需要 Python 3.10+ 與 mcp 套件，並在用戶端設定（Claude Desktop 或 Cursor）中以指令加腳本路徑啟動。未宣告任何 API 金鑰或環境變數。狀態保存在使用者家目錄下。

安裝前請注意

結論僅為建議性質：它不攔截網路流量，代理必須在付款前主動呼叫並在 BLOCKED 時停止。撤銷權杖與清除 nonce 會改動本機持久狀態。README 列出付費 Pro 方案每月 19 美元（無限稽核與威脅模擬），免費方案每月限 50 次稽核。

安裝

在 SourceWeft 中

開啟儀表板中的 Agentpay Sentinel Mcp，將其新增到工作區。
為需要使用其工具的對話啟用該服務。

Desktop only，透過 STDIO。 STDIO 服務會啟動本機處理程序，因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照儲存庫中的啟動說明。

README

[AgentPay]

AgentPay Sentinel MCP

A watchdog MCP that validates every agent payment request against all 9 security checks before it executes — catching policy violations, replay attacks, amount tampering, revoked tokens, and budget overruns at call time.

What your agent can do

Run a pre-flight audit on any payment transaction before executing it — get PASS or BLOCKED with the specific check that failed
Validate token integrity: confirms the SHA-256 hash matches the expected merchant_id:amount binding, catching forged or tampered tokens
Enforce budget caps: rejects transactions where current_spend + amount > budget_cap and returns exact remaining budget
Check merchant allowlist membership and block purchases in restricted categories
Verify token expiry and rate limits before the payment fires
Simulate known attack vectors (token forgery, budget overflow, replay, merchant spoof, expiry bypass) and get the specific defense mechanism and detection method for each

Installation

Requires: Python 3.10+, mcp package.

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

json

{  "mcpServers": {    "agentpay-sentinel": {      "command": "python",      "args": ["/absolute/path/to/agentpay-sentinel-mcp/server.py"]    }  }}

{  "mcpServers": {    "agentpay-sentinel": {      "command": "python",      "args": ["/absolute/path/to/agentpay-sentinel-mcp/server.py"]    }  }}

Cursor — add to .cursor/mcp.json in your project root:

json

{  "mcpServers": {    "agentpay-sentinel": {      "command": "python",      "args": ["/absolute/path/to/agentpay-sentinel-mcp/server.py"]    }  }}

{  "mcpServers": {    "agentpay-sentinel": {      "command": "python",      "args": ["/absolute/path/to/agentpay-sentinel-mcp/server.py"]    }  }}

Tool Reference

Tool	Description	Key params
`sentinel_audit_transaction`	Run all 9 security checks; returns PASS/BLOCKED + SHA-256 audit hash	`token_hash`, `merchant_id`, `amount`, `current_spend`, `budget_cap` (required) · `nonce`, `token_id`, `approved_amount` (enable checks 7–9)
`sentinel_revoke_token`	Permanently revoke a token — all future audits with this `token_id` will BLOCK	`token_id`, `reason`
`sentinel_clear_nonce`	Remove a nonce from the replay store (for legitimate refunds/retries only)	`nonce`
`sentinel_verify_chain`	Verify a sequence of audit hashes forms an unbroken chain	`audit_hashes`, `expected_chain_root`
`sentinel_threat_model`	Simulate any named attack vector; returns severity, defence, and detection	`attack_vector`, `context`

All 9 checks run by `sentinel_audit_transaction`

#	Check	What it catches	Param
1	Token integrity	SHA-256 hash mismatch — forged tokens	`token_hash`
2	Budget enforcement	Spend exceeding cap	`current_spend`, `budget_cap`
3	Merchant allowlist	Payment to unlisted merchant	`allowlist`
4	Category restriction	Purchase in blocked category	`blocked_categories`
5	Expiry check	Expired token	`expires_at`
6	Rate limit	Too many calls per minute	`calls_this_minute`
7	Amount mismatch	Agent changed amount after human approved	`approved_amount`
8	Replay attack	Same nonce used twice (file-backed store)	`nonce`
9	Revocation evasion	Agent using a revoked token	`token_id`

Checks 7–9 activate when the corresponding param is passed. State persists to ~/.sentinel/.

Attack vectors in `sentinel_threat_model`

token_forgery · budget_overflow · replay_attack · amount_mismatch · revocation_evasion · merchant_spoof · expiry_bypass

Security

sentinel_audit_transaction returns an advisory verdict — it does not intercept network traffic. Your agent is responsible for calling it before executing a payment and halting on BLOCKED. The audit hash returned is a SHA-256 digest of all check results, giving you a tamper-evident record of each pre-flight decision.