Privacy

This server sits between an agent and your data, so what it returns matters as much as what it redacts.

Matched values do not reach the model

The redaction tools never return the original sensitive values. The report contains filter types, character positions, confidence, and the replacement token, but not the matched text. An agent can see that an SSN was found at characters 43 to 54 and replaced, without the SSN entering its context.

The one exception is explain_redactions with include_text=true, which returns the matched values deliberately. It exists for debugging a policy against your own data locally. Do not enable it in a client connected to a hosted model.

Data stays in your infrastructure

Philter runs inside your own environment and this server is a thin client in front of it. Text sent to a tool goes to your Philter instance and nowhere else. Neither component sends data to Philterd.

Pair it with a local model

Redaction happens inside the agent's tool loop, before the model sees the redacted output. But the surrounding conversation, the prompts, the tool results, and the redacted text itself still flow to whatever model the client is configured to use.

If that model is a third-party API, anything the agent read before calling a redaction tool has already left your perimeter. Running the model locally keeps the whole loop inside your control. This is the difference between redacting data on the way to a model and redacting it after the model has already seen it.

Files are not modified

redact_file reads a file and returns redacted content. It never writes to disk. Anything that overwrites the original is a decision for the calling application, made explicitly.

Detection is probabilistic

Philter's detection is policy-driven and probabilistic. It reduces how much sensitive data passes through; it does not catch everything. Validate output against your own data, and treat the redaction report as evidence for review rather than proof of completeness.

Use explain_redactions and Philter Scope to measure a policy against representative data before relying on it.