mcpField
S
session-anonymizer

session-anonymizer

github.com/glebisSecurity401155 installsUpdated 1d ago
Repository

About session-anonymizer

session-anonymizer is a local-first skill for detecting and redacting PII in Russian and English therapy transcripts. Its sequential detection layers cover names, contact details, medications, dates, and contextual identifiers, with optional layers when dependencies are unavailable.

Key features

  1. 01Detect PII across three optional layers
  2. 02Process Russian and English transcripts
  3. 03Merge detection spans into redacted output
  4. 04Output JSON reports or pseudonyms
  5. 05Encrypt output with AES-256

Use cases

  1. 01Anonymize a single therapy transcript
  2. 02Redact multiple session files in batch
  3. 03Review detections in a JSON report

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/glebis/claude-skills --skill session-anonymizer

After installing, type/session-anonymizerin Claude Code to confirm it loaded.

Bundled filesscripts/anonymize.py

Details

From the project README

Therapy Anonymizer

Three-layer PII detection and anonymization for therapy session transcripts. Supports Russian and English. Fully local by default — no data leaves the machine.

Architecture

Three detection layers run in sequence, each catching what others miss:

LayerToolCatchesSizeSpeed
1NatashaRussian names, locations, organizations27 MBinstant
2OpenAI Privacy Filter (opf)Phones, accounts, addresses, emails2.8 GB~1.5s
3Ollama LLMMedications, dates, contextual IDs2.5-7 GB~10s

Spans from all layers are merged, overlaps resolved, and a unified redacted output is produced.

Prerequisites

pip install natasha setuptools pymorphy2-dicts-ru
pip install 'opf @ git+https://github.com/openai/privacy-filter.git'
ollama pull qwen3:4b

Each layer is optional — the script gracefully skips unavailable layers and warns.

Usage

Single file
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt
Stdin pipe
cat session.txt | python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py
Batch processing
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py --batch ~/sessions/ -o ~/sessions_clean/
JSON report
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt --json
Pseudonyms instead of tags
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt --pseudonyms
Select layers / model
# Fast — Natasha only
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt --layers natasha

# LLM only — maximum coverage
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt --layers ollama --model gemma4:e2b
Encrypt output (AES-256)
python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py session.txt -o clean.txt --encrypt "password"

Invoking from Claude Code

To anonymize text already in context, pipe it through the script:

echo '<text>' | python3 ~/.claude/skills/therapy-anonymizer/scripts/anonymize.py --json

For files, pass the path directly. Always recommend manual review after automated anonymization.

Limitations

  • Contextual identifiers ("the only red-haired architect in Kostroma") are NOT detected by any automated tool
  • OPF is English-focused — Russian coverage is partial
  • Medications detected only by Layer 3 (requires Ollama)
  • Does not assess re-identification risk from combinations of non-PII fields

Guardrails

  • NEVER send raw transcripts to cloud services
  • Cloud verification only on already-anonymized text
  • Always recommend manual review for therapy data
  • Never log original PII values

Comments

Markdown supported