# OGuardAI: Privacy Layer for LLMs

> OGuardAI is a semantic data protection runtime for AI systems. It detects sensitive entities in text, replaces them with typed semantic tokens (e.g., {{email:e_001}}), lets LLMs process the safe text, then deterministically restores originals in the output.

## Core Capabilities

- **Detect**: 16 entity types via 30+ regex patterns + GLiNER/spaCy NER
- **Tokenize**: Replace PII with semantic tokens that preserve context (gender, formality, language)
- **Transform**: Send safe text to the LLM of your choice (OpenAI, Anthropic, Mistral, local models)
- **Rehydrate**: Restore tokens to originals with 6 restore modes per output channel
- **Policy Engine**: Per-entity-type rules (tokenize, block, pass-through) with GDPR profiles
- **Output Guard**: Fail-closed second-pass scan for newly generated PII in LLM output

## Entity Types

person, email, phone, company, customer_id, order, address, location, iban, ssn, ip, url, credit_card, date_of_birth, passport, health_id

## Restore Modes

full, partial, masked, formatted, abstract, none

## Output Channels

user_output, internal_summary, customer_email, tool_payload, export, log_safe

## Architecture

- **Runtime Plane** (Rust): core engine, server, tokenizer, transformer, rehydrator, policy engine, session management
- **Integration Plane**: TypeScript SDK (@oguardai/sdk), Python SDK (oguardai-sdk), MCP server, HTTP proxy
- **Session Model**: Stateless AES-256-GCM encrypted sealed sessions with HKDF key derivation

## Deployment

Self-hosted: you run the entire runtime in your own infrastructure and hold every secret. The proprietary runtime is private today and provided under a written agreement with Oronts (request access at https://OGuard.ai/request):
- Container images: ghcr.io/oronts/oguardai/{server,detector-py,proxy} (the unit of delivery, by tag or digest)
- Helm chart (OCI): helm install oguardai oci://ghcr.io/oronts/charts/oguardai --version 0.1.0
- Docker Compose: consumes the published images
- Signed release archive: oguardai CLI + oguardai-server with SHA-256 checksum, Cosign signature, and SBOM
- Cloud recipes: ECS, Cloud Run, Container Apps (the image plus platform YAML)

The SDKs and MCP server are Apache-2.0: TypeScript (npm), Python (pip), Go (go get), Java (Maven).

## API Endpoints

- POST /v1/transform: Transform text/JSON/chat messages, return safe output + session state
- POST /v1/rehydrate: Restore tokens using session state + output channel policy
- POST /v1/detect: Detect entities only (no transformation)
- POST /v1/evaluate-policy: Evaluate policy against entities
- POST /v1/transform/stream: Streaming transform (SSE)
- POST /v1/rehydrate/stream: Streaming rehydrate (SSE)
- GET /v1/health: Health check
- GET /v1/capabilities: List entity types, languages, detectors

## Links

- Docs: https://OGuard.ai/docs
- GitHub: https://github.com/oronts/oguardai
- Company: https://oronts.com/en

## Optional

- llms-full.txt: https://OGuard.ai/llms-full.txt