ogentic-shield

Regulatory sensitivity detection for legal privilege, clinical PHI, and financial MNPI.

ogentic-shield classifies whether a piece of text — or a whole document — contains content that shouldn't leave a regulated boundary. Attorney-client privilege. HIPAA-protected clinical content. Material non-public financial information. The long tail of PII. It returns a structured AnalysisResult (score, category groups, detected entities, suggested routing) and can redact masked tokens back into plaintext after a round-trip through an external LLM.

Built for the legal, clinical, and financial AI workflows where the wrong default is "ship the prompt to OpenAI and hope."

Why this exists

On February 10, 2026, US v. Heppner (S.D.N.Y.) established that sending privileged content through a third-party AI tool can constitute waiver. The guardrail every regulated org now needs — classify before you call — didn't exist as an OSS primitive. Shield is that primitive.

Install

pip install ogentic-shield                # core (Layer 1 + 2)
pip install 'ogentic-shield[llm]'         # + Layer 3 (Ollama-backed disambiguation)
pip install 'ogentic-shield[mcp]'         # + MCP server (Claude Desktop / Goose / Cursor)
pip install 'ogentic-shield[server]'      # + FastAPI HTTP surface
pip install 'ogentic-shield[all]'         # everything

Layer 1 + 2 require the spaCy en_core_web_lg model: python -m spacy download en_core_web_lg

30-second example

from ogentic_shield import Shield

shield = Shield(profiles=["shield-legal"])

# Text-level analysis
result = shield.analyze(
    "Privileged attorney-client memo: do not disclose to opposing counsel."
)
print(result.score)                  # 0..100 sensitivity score
print(result.category_groups_found)  # {CategoryGroup.PRIVILEGE}
print(result.routing_suggestion)     # "local_only"

# Document-level redaction (v0.4.0+)
redacted = shield.redact_document("memo.txt")
print(redacted.redacted_text)        # entities replaced with deterministic tokens
print(redacted.mapping.tokens)       # token -> original, for round-trip after LLM

The full API — profiles, layers, calibration, redaction, async, MCP, HTTP server, document analysis — is documented on GitHub. See the README on GitHub for the complete reference.

What's in the box

• Three-layer detection — fast regex (Layer 1) → spaCy NER (Layer 2) → optional local-LLM disambiguation (Layer 3). Each layer adds precision without surrendering recall.
• Profile-driven — shield-legal, shield-finance, shield-healthcare, custom. Profiles define which categories to flag and at what threshold.
• Documents API — Shield.analyze_document() and Shield.redact_document() handle .txt / .md / .log today; PDF / DOCX / XLSX / EML / MSG / HTML on the roadmap.
• Token-preserving redaction — Shield.redact() and the new Shield.redact_document() substitute entities with deterministic [Label_abc123] tokens. Pair with unredact_text() to restore originals after a round-trip through OpenAI / Anthropic / Ollama.
• MCP server — ogentic-shield --mcp exposes Shield as an MCP tool surface (shield.analyze, shield.profiles, shield.calibration). Claude Desktop, Goose, and Cursor all work out of the box.
• Privacy-first — runs entirely in-process. No telemetry, no Ogentic-hosted infra. The audit row Shield emits is shape-only (hashes, scores, category names) — never the prompt text itself.

Part of a stack

ogentic-shield is the classification leg of the OgenticAI privacy-routing stack:

ogentic-shield  →  ogentic-router  →  ogentic-audit
   (classify)        (route)            (forensic log)

Together they form the open-source foundation for Sotto, OgenticAI's commercial product for regulated professionals.

Links

• Full README + API reference: https://github.com/OgenticAI/ogentic-shield#readme
• Issues / discussion: https://github.com/OgenticAI/ogentic-shield/issues
• Release notes: https://github.com/OgenticAI/ogentic-shield/releases
• Apache 2.0 license: https://github.com/OgenticAI/ogentic-shield/blob/main/LICENSE