Open-source MCP gateway blocks PII from reaching public LLMs

Hieu Luong describes a compliance problem that has emerged as AI coding assistants become standard in enterprise engineering teams: developers routinely feed real database schemas, proprietary code, and live customer records — including emails, phone numbers, and national IDs — into tools like Cursor, Windsurf, and Claude Desktop, which forward that data to public LLM API servers. This creates potential violations under GDPR, SOC 2, and local data protection regulations, while also leaving static database credentials on developer laptops and giving security operations centers no visibility into what data is leaving the perimeter.

To address these risks, the post introduces Enterprise Secure MCP Bridge, a hybrid architecture built on Anthropic's Model Context Protocol (MCP).

To address these risks, the post introduces Enterprise Secure MCP Bridge, a hybrid architecture built on Anthropic's Model Context Protocol (MCP). A local client bridge authenticates the developer against an enterprise SSO provider (such as Okta or Azure AD), receives a JWT access token, and opens an encrypted SSE connection to a central MCP Gateway hosted in the enterprise cloud. The gateway exposes only strict business-logic functions with validated schemas — rather than raw SQL tools — to limit prompt-injection risk. Before any query result is returned to the AI client, a PII redaction middleware masks emails, phone numbers, API keys, and national IDs using boundaries-aware regular expressions. All tool invocations, user identity context, input parameters, and redacted response lengths are written to an append-only `audit.log` file that can be forwarded to SIEM platforms for ISO 27001 and SOC 2 compliance.

The post reports a deployment at a 20-developer agricultural export agency where, prior to the gateway, developers were prohibited from connecting AI assistants to internal APIs at all and had to manually mock datasets. After deployment, development and debugging speeds for supply-chain APIs reportedly increased by 140%, 100% of customer and driver contact details were masked before reaching Anthropic's Claude API, and the SOC logged over 800 daily queries with full audit details and zero security incidents. The open-sourced Python boilerplate includes `gateway.py` (a FastMCP-based Starlette SSE gateway), `client_bridge.py` (a stdio-to-SSE proxy for Cursor/Windsurf), `feishu_adapter.py` (a Lark/Feishu AI adapter), and `test_integration.py` (a full integration test suite described as having 100% coverage).

Key facts

1. 01Enterprise Secure MCP Bridge is an open-sourced Python boilerplate and architecture blueprint by Hieu Luong for securing AI coding assistant connections to internal systems.
2. 02The gateway sits between AI editors (Cursor, Windsurf) and enterprise databases, authenticating developers via OAuth2/SSO (e.g., Okta, Azure AD) using JWT tokens over encrypted SSE.
3. 03Instead of exposing raw SQL tools, the MCP Gateway exposes validated business-logic functions (e.g., `search_product_inventory(sku)`) to reduce prompt-injection risk.
4. 04A PII redaction middleware masks emails, phone numbers, API keys, and national IDs before query results are returned to the AI client.
5. 05All tool invocations are logged to an append-only `audit.log` file, forwardable to SIEM platforms like Splunk or Datadog for ISO 27001 and SOC 2 compliance.
6. 06A reported deployment at a 20-developer agricultural export agency achieved a 140% speed increase in supply-chain API development and 100% masking of customer contact data sent to Claude's API.
7. 07The boilerplate includes `gateway.py`, `client_bridge.py`, `feishu_adapter.py`, and `test_integration.py` with reported 100% test coverage.

Topics