Masquerade MCP Server

Masquerade MCP is a privacy‑first solution that lets AI assistants like Claude work with sensitive documents without exposing raw data. When a user supplies only a file path, the server reads the PDF, converts it to text, and sends that content to an isolated AI platform (Tinfoil) for entity recognition. Names, emails, dates, and other personally identifiable information are automatically detected and stripped from the document before it ever reaches the LLM. The result is a fully redacted PDF that can be safely uploaded to the assistant for further analysis or question answering. By keeping the original, unredacted file out of the LLM’s view, Masquerade mitigates the risk of accidental data leakage while still enabling powerful natural‑language processing on protected content.

The server’s workflow is designed for speed and transparency. After redaction, the MCP returns a concise summary to Claude that includes masked placeholders for each sensitive field, counts of redactions per page, and the file path to the sanitized PDF. This lets developers quickly verify that all required information has been removed before any downstream processing occurs. The architecture is intentionally modular: the PDF conversion, sensitive‑data detection, and redaction steps are separated so that each can be swapped or upgraded independently. For example, the detection engine could be replaced with a different NER model without touching the rest of the pipeline.

Key capabilities include:

• File‑path based invocation – only a path is needed, keeping the original file off the LLM’s input surface.
• Automated sensitive‑data detection – leverages a dedicated, isolated AI model to identify personal and confidential information.
• Full PDF redaction – produces a new PDF with all detected entities removed, preserving layout and formatting.
• Summary reporting – provides a quick audit trail of what was redacted, facilitating compliance checks.
• Easy integration with Claude Desktop – the MCP can be registered as a tool in the assistant’s configuration, enabling seamless invocation from natural‑language prompts.

Typical use cases span legal document review, medical record anonymization, and any scenario where confidential information must be sanitized before analysis. In a contract‑review workflow, for instance, an analyst can ask Claude to “redact sensitive information from /path/to/contract.pdf,” receive a redacted file, and then proceed to ask detailed questions about clauses without risking exposure of parties’ identities. The same pattern applies to patient records, internal memos, or any PDF that contains personally identifiable data.

By acting as a privacy firewall between raw documents and AI models, Masquerade empowers developers to harness the full power of language assistants while maintaining strict data‑handling standards. Its lightweight, scriptable nature means it can be dropped into existing MCP‑enabled environments with minimal friction, making secure document processing a straightforward part of any AI workflow.