Releaseguard

ReleaseGuard

Install • Quickstart • CLI reference • Library API • Comparison • FAQ

Scan a dataset or model directory for PII and secrets with [Presidio](https://github.com/data-privacy-stack/presidio), redact what you find, and generate a public-release bundle, a Hugging Face dataset/model card plus an EU AI Act Art. 53(1)(d) training-data summary, in one command.

ReleaseGuard is not a PII detector. It's the glue between "I have a dataset I want to publish" and "I have a sanitized bundle with the paperwork already drafted." Detection is entirely Presidio's, an actively maintained open-source project with 10,000+ GitHub stars. ReleaseGuard chains Presidio's scan straight into redaction and into the two documents almost every public dataset/model release actually needs, instead of you writing a script to do it yourself.

Install

pip install releaseguard-cli

python -m spacy download en_core_web_sm # Presidio's default NLP model (~13 MB, one-time)

npm launcher (thin wrapper around the PyPI package, see "Why two registries" in FAQ)

npx releaseguard-cli --help

en_core_web_sm is spaCy's small English model and Presidio's own quickstart default. For higher-accuracy detection, install the larger model instead and pass --spacy-model:

python -m spacy download en_core_web_lg   # ~400 MB, Presidio's recommendation for production

releaseguard scan ./data --spacy-model en_core_web_lg

Quickstart

This is a real, unedited run against a two-row sample CSV, not a mockup:

$ releaseguard scan dataset --score-threshold 0.4

Scanned 1 file(s) under dataset Total findings: 7

entity type count URL 3 PERSON 2 EMAIL_ADDRESS 2

That URL: 3 line is a real Presidio quirk worth calling out rather than hiding: its regex-based URL recognizer also fires on the domain portion of an email address (example.com inside alice.rivera@example.com), so email-heavy text double-counts as both EMAIL_ADDRESS and URL. This is Presidio's own recognizer behavior, unmodified, and it's worth noting entity counts can shift slightly between Presidio/spaCy versions since detection is NLP-based, not a fixed lookup table (this run used presidio-analyzer 2.2.364 with en_core_web_sm 3.8.0). Filter it out with --entities if you only care about email addresses:

releaseguard scan dataset --entities EMAIL_ADDRESS,PERSON,PHONE_NUMBER

Redact, then package a release bundle in one command:

$ releaseguard package dataset --output bundle --score-threshold 0.4

Release bundle written to bundle dataset card: bundle/README-dataset-card.md EU AI Act Art. 53(1)(d) summary: bundle/eu-ai-act-training-summary.md redacted source: bundle-redacted-source

bundle/eu-ai-act-training-summary.md opens like this, real scan counts filled in, everything else left as an explicit placeholder for a human to complete:

# Training Data Summary (EU AI Act Art. 53(1)(d))

_Draft generated by ReleaseGuard... Based on the European Commission's training-data-summary template, published 2025-07-24. This covers the narrow, currently-binding categorical-summary requirement only -- it is not a claim of full training-data disclosure._

4. Personal Data and PII Handling (from ReleaseGuard scan)

A Presidio-backed scan (detector: presidio) covered 1 file(s) under dataset.

--json on every command switches to machine-readable output for scripts and agents.

Quick summary

• Use it for: turning a Presidio scan into a redacted copy plus a dataset/model card and an EU AI Act Art. 53(1)(d) training-data summary, from one command instead of three separate tools and a hand-written template
• What it's not: a PII detector of its own, or a claim of "total transparency"/full training-data disclosure compliance. See What ReleaseGuard is not
• Runs entirely local. No dataset content, scan results, or redacted output is ever sent to a remote service.

CLI reference

$ releaseguard --help

Usage: releaseguard [OPTIONS] COMMAND [ARGS]...

Scan, redact, and package a dataset/model directory for public release.

ReleaseGuard is a packaging layer on top of Presidio (originally a Microsoft project, now maintained by data-privacy-stack) -- it does not detect PII independently of Presidio. See the README's "What ReleaseGuard is not" section.

Options: --version Show the version and exit. --help Show this message and exit.

Commands: mcp Start an MCP server exposing scan/redact/package as agent tools. package Scan PATH, optionally redact it, and generate a release bundle... redact Scan PATH and write a redacted copy to --output. scan Scan PATH (a file or directory) for PII and secrets using...

Every command supports --json. Full flag reference: releaseguard --help.

Library API

from releaseguard.detectors import get_detector

from releaseguard.scanner import scan_directory from releaseguard.redactor import redact_directory from releaseguard.packager import build_release_bundle from releaseguard.types import RedactionStrategy

detector = get_detector("presidio", score_threshold=0.4) scan_result = scan_directory("dataset/", detector)

redaction_result = redact_directory( scan_result, "dataset-redacted/", strategy=RedactionStrategy.MASK )

bundle = build_release_bundle( scan_result, "bundle/", redaction_result=redaction_result, source_kind="dataset" ) print(bundle.eu_ai_act_summary_path)

PIIDetector (releaseguard.detectors.base) and FileReader (releaseguard.readers.base) are the two extension points. Presidio is currently the only detector shipped; CSV, JSON/JSONL, and plain text are currently the three readers shipped. Both are registries, not hardcoded calls, specifically so a new format or a second detection backend is a scoped addition later. See CONTRIBUTING.md.

MCP Server

ReleaseGuard ships a Model Context Protocol server so an AI agent (Claude, Cursor, or any MCP-compatible client) can scan, redact, and package a dataset or model directory directly, without a human invoking the CLI by hand.

Install the extra:

pip install "releaseguard-cli[mcp]"

Start it directly with the mcp subcommand:

releaseguard mcp

Add it to your MCP client's config (for Claude Desktop, claude_desktop_config.json). The server is started via a subcommand of the published releaseguard console script, not a separate console script of its own:

{

"mcpServers": { "releaseguard": { "command": "uvx", "args": ["--from", "releaseguard-cli", "releaseguard", "mcp"] } } }

Transport is stdio, so there is nothing to host: the MCP client spawns the server as a local subprocess. Source: src/releaseguard/mcp_server.py.

The server exposes three tools, each returning the same JSON shape as the matching CLI --json output:

A real example call:

scan_directory_tool(path="./dataset", score_threshold=0.4)

-> {"files_scanned": 1, "root_path": "./dataset", "findings": [...], "entity_counts": {"URL": 3, "PERSON": 2, "EMAIL_ADDRESS": 2}}

Every tool checks that path exists before doing anything else and returns a structured error, {"error": "Path ' ' does not exist.", "error_type": "PathNotFound"}, instead of a silent, misleading "0 files scanned" success result. Earlier versions of this server skipped that check and let a bad path slip straight into scan_directory, which returned a false success instead of an error; every tool handler is also now wrapped so an unexpected exception comes back as {"error": ..., "error_type": ...} rather than crashing the server. A .well-known/agent.json manifest is shipped at the repo root for A2A-style discovery, listing both the CLI and MCP interfaces and the packages that provide them.

Comparison

Star counts checked live against the GitHub API on 2026-08-03: data-privacy-stack/presidio (originally microsoft/presidio; the project moved organizations, same codebase), huggingface/huggingface_hub, [bigcode-pro