Agent Policy Gateway MCP Server

MCP server for AI agent policy enforcement. Provides PII detection, guardrails, GDPR/EU AI Act compliance checking, audit logging, and an emergency kill switch.

Features

• PII Detection — Scan text for emails, phone numbers, SSNs, credit cards, IBANs with automatic redaction
• Guardrails — Check agent actions against configurable policies (spend limits, domain whitelists, blocked actions)
• Audit Logging — Append-only JSONL audit trail per agent in ~/.agent-audit-log/
• GDPR Compliance — Check data processing actions against DSGVO requirements
• EU AI Act Compliance — Verify AI operations against EU AI Act risk categories
• Emergency Kill Switch — Instantly block an agent with critical-level audit entry

Installation

npx @aiagentkarl/agent-policy-gateway-mcp

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "policy-gateway": {
      "command": "npx",
      "args": ["-y", "@aiagentkarl/agent-policy-gateway-mcp"]
    }
  }
}

Tools

check_pii

Scan text for personally identifiable information using regex patterns.

Parameters:

• text (string) — Text to scan

Returns: Found PII types, count, masked examples, redacted text, risk level.

apply_guardrails

Check an agent action against security policies.

Parameters:

• action (string) — Action to check (e.g. "browse_url", "make_purchase")
• context (object) — Action context with optional agent_id, target_url, amount_usd, details

Returns: allow/deny/warn decision with violations and recommendations.

log_action

Write an entry to the append-only audit log.

Parameters:

• agent_id (string) — Unique agent identifier
• action (string) — Action performed
• details (string, optional) — Additional details

get_audit_log

Retrieve audit log entries for an agent.

Parameters:

• agent_id (string) — Agent identifier
• limit (number, optional) — Max entries to return (default: 50)

check_compliance

Check GDPR and EU AI Act compliance requirements.

Parameters:

• action_type (string) — Type of action (e.g. "data_collection", "profiling", "high_risk_ai", "agent_operations")
• jurisdiction (string, optional) — "gdpr", "eu_ai_act", or "both" (default)

Returns: Applicable requirements, risk level, and recommendations per framework.

emergency_stop

Kill switch — immediately block an agent from all further actions.

Parameters:

• agent_id (string) — Agent to block
• reason (string) — Reason for emergency stop

Returns: Confirmation with timestamp. Agent remains blocked until server restart.

Compliance Coverage

GDPR / DSGVO

• Data collection (Art. 6 legal basis, consent)
• Data processing (Art. 30 records, Art. 32 security)
• Data sharing (Art. 28 processor agreements)
• Data deletion (Art. 17 right to erasure)
• Profiling (Art. 22 automated decision-making)

EU AI Act

• High-risk AI systems (Art. 9-15)
• Limited-risk AI (transparency obligations)
• General-purpose AI models
• Prohibited AI practices
• Agent-specific operations

Default Guardrail Policies

| Policy | Default Value | |--------|--------------| | Max spend per action | $100 USD | | Allowed domains | github.com, stackoverflow.com, npmjs.com, pypi.org, wikipedia.org, docs.python.org, developer.mozilla.org | | Blocked actions | delete_production_data, send_mass_email, modify_permissions, access_admin_panel, execute_raw_sql, disable_security | | Require human approval | financial_transaction, data_export, user_deletion, contract_signing |

Audit Log Format

Logs are stored as JSONL files in ~/.agent-audit-log/{agent_id}.jsonl:

{"timestamp":"2026-03-20T10:30:00.000Z","agent_id":"agent-1","type":"action","action":"browse_url","details":"...","status":"logged"}

License

MIT