Skip to content

README.md

Latest commit

 

History

History
74 lines (54 loc) · 2.83 KB

README.md

File metadata and controls

74 lines (54 loc) · 2.83 KB
Raw

ShieldPrompt

Stop leaking personal data to LLMs.

pip install shieldprompt

ShieldPrompt sits between your app and any Large Language Model (ChatGPT, Claude, Gemini, a local model — doesn't matter which). Before your text leaves the machine, it swaps real emails, phone numbers, names, credit cards, and other sensitive values for reversible placeholders. When the model's answer comes back, the placeholders are swapped back so your users still see the real values.

Mask locally before sending data to a model. Review output: detection can miss PII.

The MCP helper does not intercept prompts before the model sees them. Keep originals and vaults outside model-accessible files and restore responses locally. Streaming middleware responses retain placeholders for local-client restoration.

How it works

  Your text:         "Email alice@example.com about the invoice."
                                ↓ ShieldPrompt masks
  Sent to the LLM:   "Email [EMAIL_ADDRESS_1] about the invoice."
                                ↓ LLM answers
  Model reply:       "I'll draft an email to [EMAIL_ADDRESS_1]."
                                ↓ ShieldPrompt unmasks
  User sees:         "I'll draft an email to alice@example.com."

Under the hood:

  1. Detection — a hybrid engine combines deterministic regex patterns (emails, phones, credit cards with Luhn validation, SSNs, IPs, IBANs, URLs, AWS/API keys) with Microsoft Presidio + spaCy NER for contextual entities (people, organizations, locations). If Presidio can't load, it falls back to a HuggingFace transformer model (pip install "shieldprompt[ner]").
  2. Masking — each detected value is stored in an in-memory vault and replaced with a stable token like [PERSON_1] or [EMAIL_ADDRESS_2].
  3. Unmasking — after the LLM responds, tokens are mapped back to the real values using the same vault. The vault is scoped to the request (or session) so data never leaks between users.

ShieldPrompt ships as:

  • a Python library (Shield, @mask_pii decorator),
  • a FastAPI middleware for automatic protection of chat endpoints,
  • a CLI (pii, retrace, shieldprompt) for one-off file masking,
  • an MCP server for Claude Code, Cursor, and other MCP-aware tools.

Getting started

See SETUP.md for installation, Claude Code integration, CLI usage, the Python API, and troubleshooting.

Contributing

See CONTRIBUTING.md. All changes land through pull requests reviewed by the project owner.

License

MIT