MCP Security Sanitizer

MCP Security Sanitizer

A small MCP server and interactive LangChain/Groq chatbot. The server exposes local regex-based tools for redacting common secrets and personally identifiable information (PII), and the client sanitizes each message before sending it to the language model.

Important limitation: Regex matching only recognizes the patterns

implemented here. It cannot guarantee detection of every secret or PII value.

The client guarantees that both configured sanitizer tools are called before

each non-empty user message is passed to Groq; the system prompt is

defense-in-depth, not the enforcement mechanism. Do not send regulated or

highly sensitive data without adding appropriate controls and validating

detection coverage for your environment.

Architecture

User input

| v Python client --MCP/SSE--> local regex scrubbers | 1. mask_api_keys_and_secrets | 2. mask_pii_data v Sanitized text --> LangChain agent --> Groq

Sanitization uses local Python regular expressions in the MCP server and consumes 0 LLM tokens. The client makes the sanitizer calls synchronously before invoking the agent, so raw input is not sent to Groq by this chat flow. An MCP transport is still a network boundary; this configuration binds the server to loopback (127.0.0.1) and is intended for local development.

The logging interceptor prints tool names and argument field names but withholds all argument and result contents, so it does not log unsanitized input or potentially sensitive sanitizer output.

Tools

Requirements

• Python 3.11+
• uv
• A Groq API key

Setup

From the project directory, resolve dependencies, create the project-local .venv, and write or update uv.lock:

uv sync

You do not need to activate .venv when using uv run; it selects the project environment automatically.

Create a .env file in the project directory (beside pyproject.toml):

GROQ_API_KEY=your_groq_api_key

GROQ_MODEL=openai/gpt-oss-120b

Keep .env out of version control. The client loads this file at startup and checks that GROQ_API_KEY is set before connecting to Groq. A .env placed inside .venv is also recognized as a fallback, but the project directory is the recommended location because .venv is generated and can be deleted and recreated. GROQ_MODEL is optional; if omitted, the client uses openai/gpt-oss-120b. If Groq reports that a model is unavailable, set GROQ_MODEL to a model ID enabled for your account. The API key is used by ChatGroq; it is not sent to the MCP server.

Run

Start the MCP server in one terminal:

uv run server.py

It listens on http://127.0.0.1:8000 using the SSE transport, with the MCP endpoint at /sse.

Server startup screenshot

The image below illustrates the expected server terminal after startup. It is a generated example rather than a direct screenshot from your machine; process IDs and log formatting can differ by FastMCP/Uvicorn version.

In a second terminal, start the interactive chatbot:

uv run client.py

Alternatively, activate the project environment in each terminal and run Python directly. For example, in Windows PowerShell:

.\.venv\Scripts\Activate.ps1

python server.py

In a second activated terminal:

.\.venv\Scripts\Activate.ps1

python client.py

Type exit or quit to end the chat.

Sample prompts to try

With both the server and client running, paste one of these prompts into the client. Every value below is synthetic test data; do not use real credentials or personal information for testing.

Email and phone

Write a friendly appointment reminder for alex@example.com and ask them to

call (415) 555-0134 if they need to reschedule.

The text passed to the agent should replace those values with:

alex@example.com  -> [REDACTED_EMAIL]

(415) 555-0134 -> [REDACTED_PHONE]

API key and named password

Explain why this demo credential should not be shared: sk-1234567890abcdef.

The test password is password=hunter2.

Both credential values should be replaced with [REDACTED_SECRET] before the message is sent to Groq.

Bearer token and mixed data

Help me write a support note for sam@example.org. The test phone is

212-555-0199 and the example authorization header is Bearer abc123.def456.ghi789.

The email, phone number, and Bearer token should be masked by the applicable tool. The interactive client intentionally does not print sanitized text to the terminal; it logs which tools ran while withholding their arguments and results. The expected replacements above describe what is passed onward, not what is displayed in the console.

What you see in the terminals

Run the server and client in separate terminals. The server stays running while the client connects to it.

Terminal 1 — MCP server (timestamps and process IDs vary):

INFO Starting MCP server 'MCP Security Sanitizer' with transport 'sse'

on http://127.0.0.1:8000/sse INFO: Started server process [process-id] INFO: Waiting for application startup. INFO: Application startup complete.

Terminal 2 — client (example sanitizer call; the fake values below are for testing only):

Secure assistant ready. Type 'exit' or 'quit' to stop.

You: Contact alex@example.com at (415) 555-0134; test key sk-1234567890abcdef.

[MCP] Calling mask_api_keys_and_secrets (arguments: text) [MCP] mask_api_keys_and_secrets completed; result content withheld.

[MCP] Calling mask_pii_data (arguments: text) [MCP] mask_pii_data completed; result content withheld.

Assistant: [The assistant's reply appears here after the sanitized message is sent to the configured Groq model.]

The interceptor intentionally withholds tool arguments and results from the terminal, so the unsanitized input is not printed by the MCP logs. The message passed to the agent has these substitutions:

Contact [REDACTED_EMAIL] at [REDACTED_PHONE]; test key [REDACTED_SECRET].

The assistant reply depends on the Groq model and is not a fixed output. The illustration below shows the server startup and the local sanitizer calls. It is an example terminal view, not a raw desktop screenshot; the sanitizer check does not make a Groq request. Startup timestamps and process IDs vary by run.

Configuration and production considerations

• The client requires both sanitizer tools. If either is unavailable or returns an unsupported result, the request fails instead of being sent unsanitized.
• The server uses SSE as requested. For production deployments, configure authentication, TLS, network restrictions, monitoring, and lifecycle management appropriate to your environment.
• Update the regexes and add tests against your organization's actual secret and PII formats before relying on the service for compliance.
• uv run manages and uses the project environment without requiring manual activation. To run python server.py or python client.py directly, first activate .venv in that terminal.