HALETHEIA - Ethics & Privacy by Design
Open Source Tool

xokito

Privacy engine for AI workflows. Pre-LLM deterministic obfuscation. Three privacy levels: local, obfuscated, VPC.

MIT License Python

The Privacy Paradox

AI coding assistants see everything in your codebase. API keys, customer PII, financial data. OpenAI/Anthropic promise not to train on your data. But can you afford to trust that?

# Without xokito: PII exposed
const user = { 
  email: "john.doe@acme.com",
  ssn: "123-45-6789",
  credit_card: "4532-1234-5678-9010"
};

# With xokito Level 2: Obfuscated
const user = {
  email: "TOK_EMAIL_a8f3",
  ssn: "TOK_SSN_9d2e",
  credit_card: "TOK_CC_4b7a"
};
// AI sees structure, not values. Reversible on your side.

Three Privacy Levels Explained

Choose your privacy level based on compliance requirements and infrastructure constraints

xokito Privacy Levels Comparison

🔒 Level 1: Local-First

Use case: Regulated industries (finance, healthcare, defense).

How it works: AI runs 100% locally (Ollama, LM Studio, llama.cpp). Zero external API calls. All processing happens on your hardware.

Trade-off: Smaller models (7B-70B). Less powerful than GPT-4, but mathematically zero data leakage.

Privacy: 100% | Cost: Hardware only (~$2000 GPU)

🎭 Level 2: Obfuscation (Recommended)

Use case: Most teams (balance privacy + power).

How it works: xokito intercepts code before LLM. PII replaced with deterministic tokens. "john@acme.com" → "TOK_EMAIL_a8f3". Mapping stored locally. LLM sees structure, not values.

Why deterministic? Same PII always gets same token. AI can track entities across files. "TOK_EMAIL_a8f3 in config.js is same as in auth.ts".

Privacy: 95% | Cost: Normal API rates | Reversible: Yes

🔐 Level 3: VPC Tunnel

Use case: Enterprise needing cloud LLMs with SLA.

How it works: Encrypted tunnel (TLS 1.3) to provider's VPC. Data in transit protected end-to-end. Provider contractually bound (no training on your data).

Trade-off: Trust provider's SLA. Network latency +50ms. VPC setup cost.

Privacy: 90% | Cost: +30% API premium + VPC fees

Features

🔄 Deterministic Tokens

Same PII → Same token. "john@acme.com" always becomes "TOK_EMAIL_a8f3". AI can track entities across files.

🧠 Zero Information Loss

AI sees structure + patterns. Email format preserved: "TOK_EMAIL_xxxx". Dates, IDs, patterns intact.

🔁 Reversible

You hold the key. Decode responses back to real values. AI never knows original data.

⚡ Fast

Rust-based tokenizer. 0-5ms overhead per request. You won't notice the latency.

📜 GDPR/HIPAA Ready

Compliant by design. PII never transmitted in plaintext. Audit logs included.

🔌 IDE Agnostic

Works with Cursor, Windsurf, Claude Code, or CLI. Drop-in replacement for API proxy.

Installation

<span class="text-accent"># Install via pip</span>
pip install xokito

<span class="text-accent"># Configure privacy level</span>
xokito init --level 2  # Obfuscation mode

<span class="text-accent"># Use as API proxy</span>
export OPENAI_API_BASE=http://localhost:8765
xokito start

# Your IDE now routes through xokito

Need Enterprise Features?

xokito is 100% free and open source (MIT). For enterprise teams needing centralized policy management, VPC deployments, and SLA support, check out our enterprise privacy platform.

View Xokito Enterprise →

Stop Sending PII to the Cloud

Install xokito in 5 minutes. Start protecting customer data today.

Join Waitlist PyPI Documentation
Zoomed image