mcp-compliance

MCP Audit Trail & Compliance Middleware

An MCP server that logs every AI tool call — who called it, what inputs were passed, what was returned, which model made the call, and how long it took. Built for UK FCA-regulated firms, GDPR Article 30 compliance, SOC2 Type II, and ISO 27001.

The Problem

Companies deploying Claude agents have no audit trail for MCP tool calls. Every time an agent fetches data, sends a message, or executes code — there is no log. For FCA-regulated firms (PS22/9 operational resilience), GDPR Article 30 (records of processing), SOC2, and ISO 27001, this is a compliance blocker.

What It Does

Drop this MCP server into any Claude Desktop or Claude Code setup. It exposes six tools that allow your AI agents to self-log every tool call they make:

| Tool | Description | |------|-------------| | log_tool_call | Append a tool invocation record with inputs, outputs, timing, model | | get_audit_log | Retrieve log entries filtered by date, tool, server, model | | generate_compliance_report | FCA/GDPR/SOC2/ISO27001 formatted report | | get_tool_call_stats | Aggregate metrics: calls per tool, error rates, models used | | search_audit_log | Full-text search across inputs/outputs (for DSAR requests) | | export_audit_log | CSV export for compliance submission |

Audit Log Schema

Each entry in audit.jsonl:

{
  "id": "d7a2adde-...",
  "timestamp": "2026-03-14T10:23:45.123Z",
  "session_id": "optional-session-id",
  "model": "claude-sonnet-4-6",
  "tool_server": "betfair-edge",
  "tool_name": "compute_bet_edge",
  "inputs": { "fair_prob": 0.65, "market_odds": 1.85 },
  "outputs": { "verdict": "VALUE BET", "edge_pct": 4.2 },
  "duration_ms": 234,
  "success": true,
  "error": null,
  "user_id": null,
  "tags": []
}

Stored as append-only JSONL at ~/.mcp-compliance/audit.jsonl (configurable).

Setup

1. Install

git clone <repo-url> mcp-compliance-mcp
cd mcp-compliance-mcp
npm install

2. Add to Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "mcp-compliance": {
      "command": "node",
      "args": ["/absolute/path/to/mcp-compliance-mcp/index.js"],
      "env": {
        "AUDIT_LOG_DIR": "/var/log/mcp-compliance"
      }
    }
  }
}

AUDIT_LOG_DIR is optional. Default: ~/.mcp-compliance/.

3. Add to Claude Code

{
  "mcp-compliance": {
    "command": "node",
    "args": ["/absolute/path/to/mcp-compliance-mcp/index.js"]
  }
}

Usage Patterns

Log a tool call from your agent

log_tool_call({
  "tool_server": "my-data-api",
  "tool_name": "fetch_customer_record",
  "model": "claude-sonnet-4-6",
  "user_id": "user-42",
  "inputs": { "customer_id": "C001" },
  "outputs": { "name": "Jane Smith", "tier": "premium" },
  "duration_ms": 180,
  "success": true,
  "tags": ["pii", "customer-data"]
})

Generate a monthly compliance report

generate_compliance_report({
  "start_date": "2026-03-01T00:00:00Z",
  "end_date": "2026-03-31T23:59:59Z",
  "format": "markdown"
})

Respond to a DSAR (Data Subject Access Request)

search_audit_log({
  "query": "user-42",
  "field": "all"
})

Export for auditors

export_audit_log({
  "start_date": "2026-01-01T00:00:00Z",
  "end_date": "2026-03-31T23:59:59Z"
})

Compliance Coverage

| Standard | Coverage | |----------|----------| | FCA PS22/9 | Operational resilience — AI system audit trail, third-party dependency monitoring | | GDPR Article 30 | Records of processing activities, data category inference | | SOC2 Type II | CC6 (availability), CC7 (security logging) | | ISO 27001:2022 | A.8.15 (logging), A.8.16 (monitoring) |

Configuration

| Env Variable | Default | Description | |-------------|---------|-------------| | AUDIT_LOG_DIR | ~/.mcp-compliance/ | Directory for audit.jsonl |

Architecture

index.js         — MCP server (6 tools, stdio transport)
audit-store.js   — Append-only JSONL storage with query/search/stats/CSV
report.js        — Compliance report generator (FCA/GDPR/SOC2/ISO27001)

The audit log is append-only JSONL. Each line is one JSON record. Malformed lines (e.g. from a crash mid-write) are silently skipped on read — other entries remain intact.

Requirements

• Node.js 18+
• @modelcontextprotocol/sdk ^1.27.1