Maskit is an MCP proxy that intercepts tool responses to mask sensitive data—replacing real credentials, hostnames, and PII with safe aliases before they reach AI agents.
{
"host": "prod-db.internal.net",
"api_key": "sk_live_abc123...",
"user": "john.smith@acme.com"
}{
"host": "host_1",
"api_key": "key_1",
"user": "user_42"
}MCP servers return database credentials, API keys, and internal hostnames in tool responses. Maskit intercepts these responses and replaces sensitive values with aliases before they reach the AI.
MCP database tools return real customer data—names, emails, account numbers. Maskit masks PII in tool responses so agents reason about customer_42 instead of John Smith.
GDPR, SOC2, HIPAA require that sensitive data isn't sent to third-party AI services. Maskit is the enforcement layer that sits in the data path.
MCP servers with kubectl or cloud CLI access return secrets, tokens, and internal IPs in responses. Maskit masks these values and uses guardrails to block dangerous tool calls.
Every MCP tool response goes to external AI servers. Maskit intercepts these responses and filters what the AI actually sees—real secrets stay local, only aliases are sent.
When the agent uses those aliases in future calls, Maskit transparently unmasks them back to real values before forwarding to the MCP server.
Alias sensitive values automatically based on configurable rules. Real data never reaches the AI.
Agent uses aliases naturally; Maskit swaps in real values before forwarding—zero friction.
Block tool calls that match dangerous patterns (regex, contains, equals). Prevent rm -rf disasters.
Silently inject required values (auth tokens, org IDs) the agent shouldn't see or handle.
Remove entire fields from responses. The agent never even knows they exist.
See every tool call, what was masked, what was blocked—in real-time.
stdio or HTTP transport. Zero changes to existing servers. Drop-in security.
# Coming soon
pip install maskit
maskit initMaskit is in active development. Star the repo to get notified when it's published.