PRIVACY LAB · INTERACTIVE DEMO
Keep real data out of model context.
ShadowVault replaces sensitive values with format-preserving tokens, then restores originals only at trusted tool boundaries—so schemas still work without exposing real values to model context.
Browser-only simulation · No external model call; your prompt stays in this page.
1
Detect & tokenizeFind sensitive values and substitute format-preserving tokens.
2
Model-context previewInspect the tokenized payload. This playground does not call an external model.
3
Restore at trusted toolsResolve tokens only at the tool boundary; protect tool output and streamed responses.
Session: session_playground_01
SYNTHETIC TEST FIXTURES · Example identities, payment details, and keys are fabricated for this demo. Replace them with your own non-sensitive test input.
Your agent prompt
Auto-scanned: Cards • Aadhaar • PAN • UPI • Email • Phone • SSN • Keys
Prompt scan
—
Regex & checksum engine
Trusted restore
—
@vault_guard parameter restore
Safe response stream
—
Holdback buffer per chunk
Active mappings
0 active
SQLite in-memory with TTL
01 Detect & tokenize
Detection preview
Input prompt · detected values (highlighted):
Local Tool Invocation:
stripe_chargeback_mcp_tool
Parameters received with real PII restored by @vault_guard:
Sanitized Tool Return Payload:
02 Model payload preview
Tokenized preview · simulated
Model request preview (synthetic tokens):
Downstream Schema Validations:
Streaming Re-Hydration (StreamHydrator):
03 Trusted restore + vault
In-Memory SQLite
Live Session Reverse Map:
| Type | Synthetic Token | At-Rest Storage (Vault) |
|---|
Vault Encryption:
AES-256-GCM + Blind Index
MCP Interceptor:
Stdio JSON-RPC Gateway
Privacy Policy:
Fail-Closed Enforced
TTL Expiration:
3,600s Auto-Prune