Agentguard

io.tooloracle由 tooloracle.io 發布v1.0.0更新於 Nov 2, 2026

AgentGuard — 20-tool AI safety MCP: policy preflight, risk scoring, audit logging, rate limits.

發布方自己的服務：Accessoracle

已驗證Streamable HTTP可網頁執行金融安全與監控

概覽

AI 產生的概覽

AgentGuard 是遠端 MCP 安全層，讓助理在執行工具呼叫前做政策預檢、風險評分、付款與金鑰檢查，並寫入簽章稽核日誌。

可以做什麼
  • 對預計執行的工具呼叫進行政策預檢與風險評分，回傳允許、拒絕、需核准或標記
  • 在涉及資金的動作前檢查付款政策、消費上限、速率限制與重播指紋
  • 掃描酬載與參數中的金鑰、提示注入、SQL/XSS/Shell 注入等不安全模式
  • 驗證工作階段、角色與權限範圍，並套用多租戶治理規則
  • 寫入與查詢具簽章的稽核紀錄、解釋決策結果、查詢實體威脅情報
適用情境
當助理會自主執行或涉及資金的工具呼叫，而你需要政策閘門、風險評分與可防竄改的稽核紀錄時適用，特別適合受監管、多租戶或代理商務情境。
執行需求
需要遠端 streamable HTTP 端點，可透過 mcp-remote 橋接（Node.js/npx）或直接 HTTP 連線存取；未宣告任何 API 金鑰或環境變數。需要能連線至服務商主機的網路。
安裝前請注意
工具參數會送到第三方服務掃描，酬載中的金鑰、私鑰與付款資訊會離開本機；稽核日誌由伺服器端寫入與保存。除非用戶端真的強制執行，決策僅供參考；內建政策、租戶與上限皆為服務商預設值，應自行複核。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Agentguard，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/io.tooloracle%2Fagentguard/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "agentguard": {
      "type": "http",
      "url": "https://feedoracle.io/guard-oracle/mcp/"
    }
  }
}

README

🛡️ AgentGuard MCP

Security, Policy & Audit Layer for AI Agent Tool Execution — 20 tools | Part of FeedOracle & ToolOracle

[Tools] [Status] [Backend] [Tier]

AgentGuard is the governance and security layer for AI agent workflows. Before any tool executes, AgentGuard evaluates policies, scores risk, detects secrets and injection attempts, logs to a tamper-evident audit trail, and explains every decision. Built for regulated environments, autonomous payments, and enterprise AI agent deployments.

Quick Connect

bash
# FeedOracle (compliance-focused)npx -y mcp-remote https://feedoracle.io/guard/mcp/
# ToolOracle (agent-commerce focused)npx -y mcp-remote https://tooloracle.io/guard/mcp/
json
{  "mcpServers": {    "agentguard": {      "command": "npx",      "args": ["-y", "mcp-remote", "https://feedoracle.io/guard/mcp/"]    }  }}

How It Works — The Agent Security Loop

Agent wants to call: payment_execute({amount: 5000})         │         ▼   policy_preflight()         ← Check before execution         │   risk_score: 95             ← Critical   matched: pol-001, pol-002  ← Payment + High-risk policies   decision: require_approval ← Human gate triggered         │   approval_required()        ← Register pending approval         │   [Human approves]         │   tool executes         │   audit_log_write()          ← Record outcome with signature         │   decision_explain()         ← Exportable compliance evidence

Tools — Welle 1: Core Security (7)

ToolDescription
policy_preflightPre-flight check before any tool call. Evaluates 7 policies, computes risk score, detects threats, auto-logs. Returns allowed/denied/require_approval/flagged.
tool_risk_score0-100 risk score for tool + input. Detects secrets (API keys, passwords), prompt injection, high-value amounts. eth_gas→5, payment_execute→95+.
approval_requiredCheck if tool needs human approval. Optionally registers a pending approval request with tracking URL.
audit_log_writeWrite tool execution to persistent, cryptographically-signed audit log (SQLite WAL). Call after execution to record outcome.
audit_log_queryQuery audit trail. Filter by agent, tool, decision, time range. Paginated. Returns signed entries for tamper verification.
decision_explainHuman-readable explanation of any allow/deny decision. Pass request_id for stored entry or tool_name + tool_args for fresh analysis.
rate_limit_checkCheck agent rate limits: 200/min, 5000/hr, 50000/day. Returns per-window usage with percentage.

Tools — Welle 2: Payment Controls & Safety

ToolDescription
payment_policy_checkValidate payment against policy: amount limits (>100k warns, >1M blocks), recipient denylist, supported currencies/networks, AML thresholds (>10k fiat flagged), MiCA flags.
spend_limit_checkCheck per-call/hour/day spend limits by trust level. Default: 10k/call, 50k/hr, 200k/day. Trusted: 100k/call, 500k/hr, 2M/day.
secret_exposure_checkDeep scan for 19 secret patterns: OpenAI/GitHub/AWS/Slack keys, Bearer/Basic auth, ETH private keys, Bitcoin WIF, credit cards, SSNs, emails. Returns severity + remediation.
payload_safety_check18-pattern safety scan: prompt injection, jailbreak/DAN, role hijacking, SQL (UNION/DROP/OR 1=1), XSS, Python/JS/Shell injection, path traversal, null bytes, oversized payloads.
replay_guard_checkDetect replay attacks via SHA256 fingerprint (agent+tool+args). Configurable window (default 5 min). Returns duplicate count + first/last seen.

Tools — Welle 3: Governance & Threat Intelligence (5)

ToolDescription
cross_tool_anomaly_checkDetect anomalous patterns: risky combos (wallet-recon→transfer, AML→payment), high frequency, repeated denials (policy probing), broad reconnaissance, elevated avg risk score.
scope_checkRole-based scope control. Roles: admin, compliance_officer, trader, auditor, developer, readonly. Returns has_scope, missing scope, granting roles. Logs denials.
session_validateFull session lifecycle: create (TTL + call budget), validate (increment counter), invalidate, info. Sessions carry role, scopes, tenant, expiry.
tenant_policy_checkMulti-tenant governance. Built-in tenants: default, fintech_eu (MiCA/DORA/AMLD6), defi_protocol, enterprise_read. Per-tenant blocklists, risk limits, spend caps.
threat_intel_checkEntity threat intelligence. Auto-detects ETH addresses, IPs, domains. Checks sanctions (Tornado Cash, mixers), disposable services, behavioral analysis from audit log.

Built-in Tenants

TenantMax RiskSpend/DayFrameworks
default70100,000—
fintech_eu60500,000MiCA, DORA, AMLD6
defi_protocol8010,000,000MiCA
enterprise_read300—

Built-in Roles & Scopes

RoleScopes
adminAll scopes
compliance_officeraudit:read, compliance:read, blockchain:read, security:scan
traderblockchain:read, payment:check, payment:execute, audit:read
auditoraudit:read, audit:write, compliance:read, monitor:read
developerblockchain:read, security:scan, audit:read, monitor:read
readonlyblockchain:read, audit:read

Built-in Policies (7 Default)

PolicyConditionAction
pol-001Payment/transfer toolsrequire_approval
pol-002Risk score ≥ 80require_approval
pol-003Secret/key in payloaddeny
pol-004Rate limit exceededflag
pol-005Risk score ≤ 20allow freely
pol-006Prompt injection detecteddeny
pol-007Same tool > 50 calls/60sflag

Risk Score Guide

ScoreLevelAction
0-14MinimalProceed freely
15-39LowProceed, log for audit
40-69MediumFlag and proceed with caution
70-89HighRequire human approval
90-100CriticalBlock execution

Use Cases

  • Regulated AI Workflows: MiCA/DORA compliance requires audit trails — AgentGuard provides them automatically
  • Autonomous Payments: x402 agent payments run through approval_required gate before execution
  • Multi-tenant Platforms: Rate limiting and policy scoping per agent/session
  • Security Monitoring: Real-time detection of prompt injection and secret exposure in tool arguments
  • Compliance Reporting: Export audit log with cryptographic signatures for regulatory review

Backend: SQLite WAL-Mode

Persistent, stable, no daemon required. WAL-mode supports 1000+ writes/second. Shared between feedoracle.io and tooloracle.io — one source of truth.

Links

  • 🛡️ FeedOracle: https://feedoracle.io/guard/mcp/
  • 🔧 ToolOracle: https://tooloracle.io/guard/mcp/
  • 📚 Docs: feedoracle.io
  • 🏠 Platform: tooloracle.io

Built by FeedOracle — Evidence by Design

來源：README.md，提交 9cd9bd6

工具

0
SourceWeft 尚未讀取此伺服器的工具。

版本歷史

1
  1. v1.0.0最新Sep 16, 2026

評分

已滿足適用檢查項的 75%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分
  • 已宣告認證方式必需10 分
  • 已列出工具必需15 分

    請讓伺服器在無憑證時回應 tools/list，或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

    請在儲存庫或註冊中心條目中宣告授權條款。

其他 Tooloracle 整合

Accessoracle

io.tooloracle官方發布者

AccessOracle 是遠端 MCP 伺服器，提供 10 個工具，用於追蹤 IAM 與特權存取合規，包括 MFA 涵蓋率、存取複審、職責分離衝突等。

Streamable HTTP可網頁執行
安全與監控10 個工具Oct 24, 2026

Zkevidenceoracle

io.tooloracle官方發布者

一個遠端 MCP 伺服器，提供 14 個用於合規證據的零知識證明工具，包括 Groth16 與 PLONK。

Streamable HTTP可網頁執行
安全與監控尚未收錄Oct 24, 2026

Yieldoracle

io.tooloracle官方發布者

遠端 DeFi 收益情報伺服器，讓助理存取包含風險調整後 APY 與 RWA 收益資料的大型池子資料集。

Streamable HTTP可網頁執行
金融尚未收錄Oct 24, 2026

Trustlayer

io.tooloracle官方發布者

一個遠端 MCP 伺服器，提供十個跨伺服器信任工具，用於簽章、錨定與驗證。

Streamable HTTP可網頁執行
安全與監控尚未收錄Oct 24, 2026

Tonoracle

io.tooloracle官方發布者

遠端 MCP 伺服器，為助理提供 11 個唯讀 TON 區塊鏈工具，涵蓋代幣、錢包、DeFi 收益、協議、穩定幣、跨鏈橋和網路統計。

Streamable HTTP可網頁執行
金融尚未收錄Oct 24, 2026