shutapp 3.0.0

Model-agnostic PII firewall for LLM apps: scan/redact/rescan/verify SDK + a local Claude Code redaction proxy (`shutapp claude`).

shutapp — the model-agnostic PII firewall

Three lines. Any LLM.

from shutapp import Shutapp

g = Shutapp(api_key="shu_live_...")
safe, session = g.scan("My email is [e-mail]")

# ...send `safe` to any LLM using your own provider key...

final = g.unscan(response_text, session)

Or the 2-line auto-wrap for OpenAI / Anthropic:

from openai import OpenAI
from shutapp import guard
client = guard(OpenAI())          # every .chat.completions.create call is now protected

from anthropic import Anthropic
from shutapp import guard
client = guard(Anthropic())       # every .messages.create call is now protected

Run Claude Code privately — shutapp claude

The same package ships a local CLI that wraps Claude Code so secrets are redacted before they ever reach Anthropic:

pip install shutapp
shutapp claude            # launches Claude Code through a local redaction proxy

shutapp claude starts a localhost proxy and runs the real claude through it, so API keys and other secrets are swapped for reversible placeholders on the way out and restored locally — Anthropic only ever sees redacted text. It works with both API-key and Claude Max/Pro subscription logins: on a subscription it auto-bridges your OAuth through the proxy (your credentials never leave your machine except to Anthropic, exactly as Claude Code already sends them). Requires Claude Code's claude on your PATH.

Local chat history is redacted too

The model never receives a raw secret: the proxy re-redacts the entire conversation (including replayed history) on every API call. But Claude Code also writes your raw typed prompt to its own local transcript files (~/.claude/projects/.../*.jsonl) for --resume. shutapp claude runs a real-time history redactor that rewrites those secrets to the same placeholder the model saw, and does a guaranteed final pass when Claude exits. Opt out with shutapp claude --no-history-scrub (the model stays protected either way).

Already have older transcripts with secrets in them? Clean them up:

shutapp history scrub --dry-run   # report only
shutapp history scrub             # redact transcripts from the last 7 days
shutapp history scrub --all       # redact every transcript

shutapp status            # version, vault dir, detector count
shutapp session list      # value-free session stats
shutapp detectors list    # the bundled detector registry
shutapp verify rcpt.json  # offline receipt verification

What it does

1. Sends your prompt text to the Shutapp API, which replaces PII with reversible placeholders like [EMAIL_ADDRESS_1], [PHONE_NUMBER_1], [API_KEY_1].
2. You send the redacted text to any LLM using your own key. Shutapp never sees your LLM key and never sees the response.
3. You call unscan(...) to restore the original values in the LLM's response.

After every scan Shutapp prints a one-line receipt to your terminal:

[shutapp] scan ok · 12ms · EMAIL_ADDRESS x1, PHONE_NUMBER x1 · rcpt_8f3a2b

Install (from PyPI, later)

pip install shutapp

Install (from this monorepo, right now)

pip install -e ./packages/sdk-python

Config

Shutapp(api_key, base_url, receipts, receipt_format, timeout)

• receipts: "stderr" (default), "stdout", "silent", or a callable fn(receipt_dict) for custom logging.
• receipt_format: "line" or "block".
• base_url: override for self-hosted Shutapp (e.g. http://localhost:8080).