Product
Everything between your data and the model
Detection, policy, surrogates, an encrypted vault and rehydration, delivered as a REST API, an LLM gateway and an MCP proxy.
Open-source detection you can inspect
Built on the Presidio analyzer framework and spaCy, with deterministic validators for anything that has a checksum. Coreference clustering makes every mention of the same person or organization map to one surrogate.
- PERSON, ORGANIZATION, LOCATION, ADDRESS, DATE
- EMAIL, PHONE, URL, IP_ADDRESS
- SSN, CREDIT_CARD (Luhn), IBAN (mod-97), ACCOUNT_NUMBER, MEDICAL_ID
- Tenant dictionaries and allow-lists
- Optional GLiNER model and LLM verifier for difficult text
{
"entities": [
{"type": "PERSON", "start": 6, "end": 14, "score": 0.93,
"source": "spacy", "cluster_id": "c1"},
{"type": "EMAIL", "start": 18, "end": 30, "score": 1.0,
"source": "checksum"}
],
"detectors": ["presidio", "spacy", "checksum", "dictionary"]
}Realistic by default
Models reason better over natural text. Zanitize generates locale- and gender-consistent names and format-preserving values, and shifts dates consistently so intervals still make sense.
- Names that keep gender and locale
- Emails, phones, cards and IBANs that keep their format
- Consistent date shifting per conversation
- Token mode (<PERSON_1>) and hashing when you prefer opaque values
- Deterministic HMAC mode for stable tenant-wide pseudonyms
{
"text": "Email Maria Lopez at maria.lopez@northwind.example",
"session_id": "ses_123",
"counts": {"PERSON": 1, "EMAIL": 1},
"action_counts": {"pseudonymize": 2}
}An OpenAI- and Anthropic-compatible proxy
Point your SDK's base URL at Zanitize. Messages, system prompts and tool results are anonymized on the way out; text deltas and tool-call arguments are rehydrated on the way back, including SSE streams.
- Chat Completions, Responses, Embeddings and Messages APIs
- Streaming rehydration that never emits half a surrogate
- Tool calls: arguments rehydrated, results anonymized
- Multi-turn conversations share one vault session
- Destination-model conditions in policy (e.g. block SSNs to external providers)
from openai import OpenAI
client = OpenAI(
base_url="https://api.zanitize.io/v1/proxy/openai/v1",
default_headers={"X-API-Key": "zk_live_..."},
)
client.chat.completions.create(model="gpt-4o", messages=[...])Originals never leave the vault
Mappings are stored with per-tenant envelope encryption. Choose how long they live and at what scope, and destroy them for good by shredding the data key.
- Local master key or AWS KMS, including bring-your-own-key
- Request, conversation or tenant scope with TTLs
- Crypto-shredding per conversation or tenant
- Logs and audit events carry counts only, never values
{"shredded": 14}Policies, audit and compliance
Policies are versioned YAML you can validate, dry-run and roll out in shadow mode. Every request emits a hash-chained audit event, and compliance reports map controls to your configuration.
- Validate and simulate before you activate
- Shadow mode: observe what would be blocked without changing traffic
- Audit log export (CSV, JSONL) with chain verification
- HIPAA, GDPR and SOC 2 evidence reports
- RBAC, OIDC single sign-on and scoped API keys
{"valid": true, "events_checked": 18234, "first_bad_event_id": null}Themes for demos and the free tier
Production traffic uses realistic surrogates. For workshops, demos and teaching, you can switch a request to a themed cast instead: original characters we created, or public-domain folklore such as Japanese yokai, Journey to the West and Greek myth. The available themes come straight from the API.
Try Fun modeX-Zanitize-Surrogates: universe
X-Zanitize-Universe: <theme id>