mcp-eu-comply

mcp-eu-comply

August 2, 2026. €35M fines. Is your MCP server ready?

mcp-eu-comply is the first runtime EU AI Act compliance wrapper for MCP servers.

One function call adds tamper-evident audit logging, human-in-the-loop oversight, risk classification, and PII redaction to any MCP server — designed to meet EU AI Act Article 12, 14, and 19 requirements.

What it does

• Audit logging with SHA-256 hash chain — tamper-evident NDJSON logs (Article 12)
• Human-in-the-loop oversight — pause, approve, or deny tool calls via webhook or custom handler (Article 14)
• Risk classification — pattern-based, per-tool risk levels aligned with EU AI Act categories (Article 9)
• PII redaction — deep recursive field redaction before storage (GDPR Article 5)
• Compliance report generation — JSON summaries for auditors, covering any time period

Try it in 10 seconds

npx mcp-eu-comply demo

This runs a self-contained simulation: 7 tool calls across 4 risk levels, PII redaction, human oversight (approve/deny/timeout), and hash chain verification. No config needed.

Use --keep to save the generated audit logs for inspection:

npx mcp-eu-comply demo --keep

Quick Start

npm install mcp-eu-comply

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { wrapWithCompliance } from "mcp-eu-comply";

const server = new McpServer({ name: "my-server", version: "1.0.0" });

const compliantServer = wrapWithCompliance(server, {
  riskRules: [
    { toolPattern: /delete|drop|remove/i, level: "critical" },
    { toolPattern: /write|update|send/i, level: "high" },
    { toolPattern: /.*/,                  level: "medium" },
  ],
  logging: {
    outputDir: "./audit-logs",
    retention: { days: 365 },
  },
  oversight: {
    requireApproval: ["critical"],
    notifyOn: ["high"],
    webhook: "https://your-company.eu/oversight",
    timeoutMs: 30_000,
    onTimeout: "deny",
  },
  dataResidency: {
    region: "EU",
    piiFields: ["email", "name", "address", "phone", "iban"],
    redactInLogs: true,
  },
});

// Register tools on compliantServer — they are automatically wrapped
compliantServer.tool("transfer_funds", { amount: {}, to: {} }, async (args) => {
  // Your logic here. mcp-eu-comply handles the rest.
  return { content: [{ type: "text", text: "Done" }] };
});

How it works

[Agent] → [MCP Tool Call] → [mcp-eu-comply Proxy]
                                      │
                             Risk Classification
                             (pattern-match tool name + args)
                                      │
                             Human Oversight Check
                             (webhook or custom handler)
                                      │
                             Original Tool Executes
                             (untouched — proxy is transparent)
                                      │
                             Audit Log (SHA-256 hash chain)
                             (PII redacted, NDJSON, append-only)

The wrapper uses a JavaScript Proxy on the McpServer instance. It intercepts tool and registerTool calls to wrap each callback. All non-intercepted methods pass through via Reflect.get. Errors in the compliance layer are caught and logged — they never break tool execution.

Example audit entry

Each line in the NDJSON log file is a self-contained JSON object:

{"id":"a1b2c3d4-5678-4ef0-abcd-1234567890ab","timestamp":"2026-07-15T14:32:01.442Z","prevHash":"e3b0c44298fc1c149afbf4c8996fb924","hash":"9f86d081884c7d659a2feaa0c55ad015","tool":"transfer_funds","args":{"amount":500,"to":"***REDACTED***","from":"***REDACTED***"},"risk":"critical","oversight":{"required":true,"status":"approved","approvedBy":"[e-mail]","approvedAt":"2026-07-15T14:32:00.112Z"},"result":{"status":"success","contentHash":"b94d27b9934d3e08a52e52d7da7dabfac484efe37a5380ee9088f7ace2efcde9"},"durationMs":1843,"agentId":"agent-47","sessionId":"sess-001","schemaVersion":"0.1.0"}

Key fields: prevHash and hash form a tamper-evident chain. args are PII-redacted. result.contentHash stores a SHA-256 digest — never the raw output.

Configuration reference

Option	Type	Default	Description
riskRules	RiskRule[]	[]	Pattern-based rules mapping tools to risk levels
riskRules[].toolPattern	RegExp | string	—	Pattern to match against tool name
riskRules[].level	RiskLevel	—	Risk level when matched
riskRules[].argsPattern	Record<string, RegExp | string>	—	Optional args matching
logging.outputDir	string	—	Directory for NDJSON logs and chain state
logging.retention.days	number	—	Log retention in days
logging.hashAlgorithm	'sha256' | 'sha384' | 'sha512'	'sha256'	Hash algorithm for the chain
oversight.requireApproval	RiskLevel[]	[]	Risk levels requiring human approval
oversight.notifyOn	RiskLevel[]	[]	Risk levels that trigger notifications
oversight.webhook	string	—	Webhook URL for approval requests
oversight.handler	OversightHandler	—	Custom handler (alternative to webhook)
oversight.timeoutMs	number	—	Max wait time for human response (ms)
oversight.onTimeout	TimeoutAction	'deny'	Action on timeout: deny, allow, or escalate
dataResidency.region	DataRegion	—	Data residency region (EU, FR, DE, custom)
dataResidency.piiFields	string[]	[]	Field names to redact (case-insensitive)
dataResidency.redactInLogs	boolean	false	Enable PII redaction in audit logs

Risk levels

Level	EU AI Act Category	When to use	Default behavior
low	Minimal risk	Read-only lookups, status checks	Log only
medium	Limited risk	Data writes, standard operations	Log only (default when no rule matches)
high	High risk	Financial ops, PII access, bulk changes	Log + notify
critical	Unacceptable risk threshold	Deletions, privilege changes, fund transfers	Log + require human approval

Oversight configuration

Webhook mode

POST requests are sent to your webhook URL with an OversightRequest body. Respond with an OversightDecision:

// Your webhook receives:
{
  id: "req-uuid",
  tool: "transfer_funds",
  args: { amount: 500, to: "***REDACTED***" },
  risk: "critical",
  timestamp: "2026-07-15T14:32:00.000Z",
  context: { sessionId: "sess-001", agentId: "agent-47" }
}

// Your webhook responds:
{
  status: "approved",       // or "denied"
  approvedBy: "[e-mail]",
  reason: "Transfer under €1000 threshold"
}

Custom handler mode

For Slack bots, internal dashboards, or queue-based flows:

import type { OversightHandler } from "mcp-eu-comply";

const slackHandler: OversightHandler = {
  async requestApproval(request) {
    // Post to Slack, wait for button click
    const decision = await postToSlackAndWait(request);
    return {
      status: decision.approved ? "approved" : "denied",
      approvedBy: decision.user,
      reason: decision.comment,
    };
  },
  async notify(notification) {
    await postToSlackChannel(`Tool ${notification.tool} called (risk: ${notification.risk})`);
  },
};

const server = wrapWithCompliance(mcpServer, {
  // ...
  oversight: {
    requireApproval: ["critical", "high"],
    handler: slackHandler,
    timeoutMs: 60_000,
    onTimeout: "deny",
  },
});

Regulatory coverage

Regulation	Articles	What mcp-eu-comply covers
EU AI Act	12, 14, 19	Tamper-evident logging, human oversight engine, structured log quality
GDPR	5, 17, 25	PII redaction, data minimisation, retention policy, EU residency tagging
DORA	11, 12	Complete audit trails for ICT incident analysis, backup-ready NDJSON

Verify chain integrity

The hash chain can be verified at any time to detect tampering:

import { verifyChain } from "mcp-eu-comply";

const result = await verifyChain("./audit-logs");

console.log(result);
// { valid: true, entries: 14208 }

// If tampered:
// { valid: false, entries: 14208, firstBrokenAt: 9451, error: "Hash mismatch at entry 9451" }

CLI Tools

Run an interactive demo:

npx mcp-eu-comply demo
npx mcp-eu-comply demo --keep              # Save generated audit logs

Verify your audit chain:

npx mcp-eu-comply verify --dir ./audit-logs

Generate a compliance report:

npx mcp-eu-comply report --dir ./audit-logs --format human

Filter by agent in multi-agent setups:

npx mcp-eu-comply verify --dir ./audit-logs --agent payment-service

Exit codes: 0 = valid chain, 1 = broken chain, 2 = file error.

Templates

Pre-built compliance configurations for regulated industries:

import { wrapWithCompliance, doraFintech } from "mcp-eu-comply";

const server = wrapWithCompliance(mcpServer, {
  ...doraFintech,
  logging: { outputDir: "./audit-logs" },
});

Available templates:

Template	Industry	Risk rules	Oversight	PII fields	Retention
doraFintech	Financial services (DORA)	payment/transfer → critical	critical + high require approval	10 fields incl. IBAN, BIC	1825 days (5 years)
gdprEcommerce	E-commerce (GDPR)	delete/drop → critical	critical requires approval	6 fields incl. credit_card	—

Multi-Agent Support

For setups with multiple MCP servers sharing the same log directory:

const server = wrapWithCompliance(mcpServer, {
  ...config,
  agentId: "payment-service",  // Each agent gets its own hash chain
});

Each agent maintains an independent hash chain (chain-state-{agentId}.json). Entries are stored in shared NDJSON files with the agentId field. Verify or report on a single agent's chain:

npx mcp-eu-comply verify --dir ./audit-logs --agent payment-service

Important disclaimer

This package is designed to meet EU AI Act Article 12, 14, and 19 requirements. It is NOT certified or officially approved — the CEN/CENELEC harmonised standards are still being drafted. Use as part of a broader compliance strategy.

Zero external runtime dependencies. 87 tests. Peer dependency on @modelcontextprotocol/sdk >= 1.0.0. Node.js >= 18.

Roadmap

• SHA-256 hash chain audit logging
• Human oversight engine (webhook + custom handler)
• Pattern-based risk classification
• Deep recursive PII redaction
• Compliance report generation
• CLI validator (npx mcp-eu-comply verify + report)
• Interactive demo (npx mcp-eu-comply demo)
• PDF audit reports
• Dashboard SaaS
• DORA fintech + GDPR e-commerce templates
• Multi-agent chain isolation
• eIDAS 2.0 identity bridge

License

MIT