The drop-in AI privacy gateway that tokenizes sensitive data before any prompt leaves your organization — then reverses it on the way back. Role-aware. Audit-logged. Built for regulated finance.
Prevent GLBA violations · Satisfy GDPR · Preserve CCPA rights · Deploy in under 30 minutes
Product
Both. Vaultex is a gateway that fits into your existing workflow — no matter how you work with LLMs.
Point any Anthropic, OpenAI, or Ollama SDK at the Vaultex gateway URL. Every outbound prompt is auto-tokenized. Every response is auto-detokenized. Zero refactoring.
# Before
client = anthropic.Anthropic(api_key="sk-ant-...")
# After — one change
client = anthropic.Anthropic(
base_url="https://your-vaultex-host:8000",
api_key="sk-ant-..."
)Open the built-in web interface at /chat. Upload a CSV of customer records, start a conversation, and watch PII get masked in real time — role-filtered per your RBAC config.
Drop a CSV with thousands of loan records into the /chat interface. Ask portfolio-level questions. The LLM sees tokens, computes real analytics, and you see role-filtered answers.
The Problem
Financial institutions face compounding pressure to adopt AI — but every unguarded prompt is a potential compliance incident.
Every time a team member copies customer data into an AI prompt, your org risks a GLBA violation with no log of what left your network or which model processed it.
Cloud LLMs process and may retain prompt data. Without a proxy layer, you have zero regulator-ready evidence of what sensitive information reached which external model.
Most AI tools have no row-level access control. A junior analyst sees the exact same AI response as a VP — including full PII they should never be authorized to access.
Capabilities
Everything your team needs to safely connect to any LLM without exposing customer data, regulated PII, or financial records.
Microsoft Presidio NER catches PERSON, SSN, ACCOUNT_NUMBER, LOAN_ID, EMAIL, PHONE, and DATE_OF_BIRTH before any prompt crosses your network boundary.
PII becomes {{PERSON_1}} — not random noise. The same person gets the same token in every message. Responses auto-detokenize on the way back.
Balances, credit scores, rates, and risk flags are NEVER masked. Your LLM can still compute averages, flag high-risk loans, and run full distributions.
Route to Anthropic Claude, OpenAI GPT-4, or self-hosted Ollama. The gateway tokenizes PII locally before any request reaches a cloud API — so neither Anthropic nor OpenAI ever sees a real name, SSN, or account number.
Junior analysts see {{PERSON_1}}. Senior analysts see the name. VPs see everything. RBAC enforced at the token level, not the application layer.
Every request, PII detection event, and auth failure is logged with correlation IDs. 30-day retention. Regulator-ready admin console included.
Integration
Vaultex is a drop-in proxy. Point your existing SDK at our gateway endpoint and every prompt is automatically tokenized — no code changes, no schema migrations, no training required.
# Before Vaultex
import anthropic
client = anthropic.Anthropic(
api_key="sk-ant-..."
)
# After Vaultex — one change
client = anthropic.Anthropic(
base_url="https://your-gateway-host:8000",
api_key="sk-ant-..."
)
# Every prompt is now auto-tokenized.
# Zero code changes beyond base_url.Banking Analytics
Vaultex only masks direct personal identifiers. Financial dimensions — the data your analysts actually need — flow through untouched. Your LLM can compute averages, distributions, and risk clusters without ever seeing a real name or SSN.
How It Works
Choose Anthropic, OpenAI, or a local Ollama model. Enter your API key — encrypted at rest with Fernet. Point your SDK at the Vaultex base URL. Done in 30 seconds.
Upload a CSV with customer records. The gateway intercepts every prompt, tokenizes PII, and forwards the sanitized version to your chosen LLM. Analytics fields pass through intact.
Switch personas live. Watch how a junior analyst sees tokens while a VP sees real names. RBAC enforced at the token layer — no application code changes needed.
Why Vaultex
Generic DLP tools scan files. Vaultex tokenizes live AI prompts — and preserves analytics. That is a different product entirely.
~ = partial / varies by vendor implementation
Access Control
Role-Based Decryption — enforced at the token layer, not the application layer.
Watch a live banking risk query flow through the Vaultex gateway — tokenized on the way out, detokenized on the way back, fully audit-logged throughout.
Simulated walkthrough · Real gateway uses Microsoft Presidio NER with 14+ entity types · avg +12ms overhead
Try It Now
Edit the sample text, then click Tokenize. PII is detected client-side in this preview — the real gateway uses Microsoft Presidio NER with 14+ entity types.
Pricing
The core tokenization engine is MIT open-source and always will be.
We earn by selling the features regulated institutions actually need: cloud LLM routing, SOC 2 reports, GLBA evidence packs, SSO, extended retention, and dedicated SLAs. Open source builds trust; enterprise features fund the team.
Self-host with Ollama. Full tokenization engine, no credit card, no time limit.
Unlock Anthropic + OpenAI routing, full compliance tooling, and team access.
For regulated institutions — on-prem VPC, SOC 2, GLBA evidence packs, and SLAs.
All plans include end-to-end encryption, RBAC, and audit logging. Compare all features in detail
FAQ
Early Access
Be first to access multi-LLM routing, 90-day audit retention, and role-based export controls. Early access members receive 3 months free.