The privacy firewall for AI apps and agents
Zanitize finds names, emails, account numbers and health data in prompts, tool calls and files, swaps them for realistic, consistent surrogates before they reach a model, and restores the originals in the response. Your models stay useful. Your data stays yours.
No credit card. Drop-in for OpenAI, Anthropic and any MCP server.
Your app sends
Draft a reply to Priya Natarajan (priya.n@harborhealth.org) about claim HX-448-2210. Her card ending 4242 was charged twice on March 3.
The model sees
Draft a reply to Elena Marsh (elena.marsh@northwind.example) about claim KQ-917-5536. Her card ending 8801 was charged twice on March 17.
Your app receives
Hi Priya, I'm sorry about the duplicate charge on March 3. I've escalated claim HX-448-2210 and a refund to the card ending 4242...
Why Zanitize
Redaction breaks your prompts. Surrogates don't.
Masking every name as [REDACTED] confuses models and loses who-did-what. Zanitize keeps the text natural and the answer reversible.
Accurate detection
Presidio and spaCy NER, checksum-validated cards, IBANs and SSNs, your own dictionaries, plus optional GLiNER and an LLM verifier for hard cases.
Realistic, consistent surrogates
Locale- and gender-consistent names, format-preserving emails, phones, cards and addresses, and consistent date shifting. The same person is the same surrogate all conversation.
Streaming rehydration
Originals are restored in responses, even in SSE streams where a surrogate is split across chunks. Possessives, plurals and partial names are handled.
Policy as code
Versioned YAML per tenant, app, route and tool: allow, mask, pseudonymize, tokenize, generalize, hash or block, with JSONPath targeting and shadow mode.
Encrypted vault
Per-tenant envelope encryption with a local master key or AWS KMS (BYOK), TTLs, request/conversation/tenant scopes and crypto-shredding.
Drop-in gateway
OpenAI- and Anthropic-compatible endpoints with streaming and tool calls. Change a base URL, keep your SDK.
Built for agents
Put a firewall between your agent and its tools
Agents read CRMs, inboxes and databases through MCP tools, then pass everything to a model. The Zanitize MCP proxy sits in the middle: tool results are anonymized before the model sees them, and tool-call arguments are rehydrated so your tools still receive real values.
- Per-tool policies:
crm.*outputs pseudonymized,payments.*blocked - JSONPath field targeting for structured tool results
- One conversation, one consistent set of surrogates across every turn
- Config generator for Claude Desktop, Cursor and any Streamable HTTP client
id: support-agent
default_action: pseudonymize
scope: conversation
surrogate_mode: realistic
allow_list: ["Acme Support"]
rules:
- id: mask-cards
entity_types: [CREDIT_CARD]
action: mask
params: {keep_last: 4}
- id: block-ssn-external
entity_types: [SSN]
destination_providers: ["openai"]
action: block
- id: crm-results
tools: ["crm.*"]
tool_direction: output
fields: ["$.records[*].notes"]
action: pseudonymizeHow it works
Every request takes the same auditable path, in milliseconds.
- 1
Detect
Find PII and secrets in text, JSON, tool results and files.
- 2
Decide
Your policy picks an action per entity, field and tool.
- 3
Substitute
Realistic surrogates replace the originals; mappings go to the vault.
- 4
Call the model
Only surrogates leave your boundary.
- 5
Rehydrate
Originals are restored in the response and tool arguments.
Evidence your auditors will accept
Append-only, hash-chained audit logs record what was protected and by which policy, without ever storing the values themselves.
Security overviewHIPAA
PHI de-identification controls mapped to your policies, with evidence reports.
GDPR
Pseudonymization (Art. 4(5)), EU data residency and crypto-shredding for erasure.
SOC 2
RBAC, SSO, scoped keys and a verifiable audit trail for change and access controls.
Ship AI features without shipping your customers' data
Start on the free plan, point your SDK at the gateway, and watch the audit log fill with counts, not secrets.