🛡️ AgentGuard MCP

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

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/

config.json

{
  "mcpServers": {
    "agentguard": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://feedoracle.io/guard/mcp/"]
    }
  }
}

How It Works — The Agent Security Loop

Code

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)

Tool	Description
`policy_preflight`	Pre-flight check before any tool call. Evaluates 7 policies, computes risk score, detects threats, auto-logs. Returns `allowed`/`denied`/`require_approval`/`flagged`.
`tool_risk_score`	0-100 risk score for tool + input. Detects secrets (API keys, passwords), prompt injection, high-value amounts. `eth_gas`→5, `payment_execute`→95+.
`approval_required`	Check if tool needs human approval. Optionally registers a pending approval request with tracking URL.
`audit_log_write`	Write tool execution to persistent, cryptographically-signed audit log (SQLite WAL). Call after execution to record outcome.
`audit_log_query`	Query audit trail. Filter by agent, tool, decision, time range. Paginated. Returns signed entries for tamper verification.
`decision_explain`	Human-readable explanation of any allow/deny decision. Pass `request_id` for stored entry or `tool_name` + `tool_args` for fresh analysis.
`rate_limit_check`	Check agent rate limits: 200/min, 5000/hr, 50000/day. Returns per-window usage with percentage.

Tools — Welle 2: Payment Controls & Safety

Tool	Description
`payment_policy_check`	Validate payment against policy: amount limits (>100k warns, >1M blocks), recipient denylist, supported currencies/networks, AML thresholds (>10k fiat flagged), MiCA flags.
`spend_limit_check`	Check 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_check`	Deep 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_check`	18-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_check`	Detect 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)

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

Built-in Tenants

Tenant	Max Risk	Spend/Day	Frameworks
`default`	70	100,000	—
`fintech_eu`	60	500,000	MiCA, DORA, AMLD6
`defi_protocol`	80	10,000,000	MiCA
`enterprise_read`	30	0	—

Built-in Roles & Scopes

Role	Scopes
`admin`	All scopes
`compliance_officer`	audit:read, compliance:read, blockchain:read, security:scan
`trader`	blockchain:read, payment:check, payment:execute, audit:read
`auditor`	audit:read, audit:write, compliance:read, monitor:read
`developer`	blockchain:read, security:scan, audit:read, monitor:read
`readonly`	blockchain:read, audit:read

Built-in Policies (7 Default)

Policy	Condition	Action
pol-001	Payment/transfer tools	`require_approval`
pol-002	Risk score ≥ 80	`require_approval`
pol-003	Secret/key in payload	`deny`
pol-004	Rate limit exceeded	`flag`
pol-005	Risk score ≤ 20	`allow` freely
pol-006	Prompt injection detected	`deny`
pol-007	Same tool > 50 calls/60s	`flag`

Risk Score Guide

Score	Level	Action
0-14	Minimal	Proceed freely
15-39	Low	Proceed, log for audit
40-69	Medium	Flag and proceed with caution
70-89	High	Require human approval
90-100	Critical	Block 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.

Agentguard README