By what you are building
| You want to | Use | Shape |
|---|---|---|
| Redact inside your own code | Phileas | Library (Java, Python, .NET) |
| Call redaction over HTTP | Philter | Self-hosted service |
| Redact prompts before they reach an LLM provider | Philter AI Proxy | Drop-in proxy |
| Redact from Claude Code, Claude Desktop, or Cursor | Philter MCP | MCP server, needs Philter |
| Find where sensitive data already lives | Phinder | Scanner |
| Put a human in the loop on detections | Arbiter | Review application |
| Measure a policy on precision and recall | Philter Scope | Evaluation tool |
| Alert when PII volumes change | Phield | Monitor |
| Publish aggregate counts safely | Philter Diffuse | Differential privacy |
| Build policies without writing JSON | Redaction Policy Editor | Web console |
| Express policies as queries | PhiSQL | Query language |
Library or service
This is the decision most people actually face.
Phileas runs in your process. No network hop, no deployment, no operational surface. Choose it when redaction belongs inside an existing application and you are working in Java, Python, or .NET.
Philter wraps the same engine in a service with an HTTP API, a dashboard, policy management, and clustering. Choose it when several applications share one policy set, when the callers are not JVM or Python programs, or when you want redaction managed and audited separately from the applications using it.
Both are self-hosted. Neither sends your data anywhere.
If you are working with LLMs
Philter AI Proxy sits between your application and the provider (OpenAI, Anthropic, Bedrock, Gemini, Ollama, and OpenAI-compatible endpoints). Prompts are redacted on the way out and responses are scanned on the way back. Your application keeps talking to the same API.
Philter MCP exposes redaction as tools an MCP client can call, so an agent can redact inside its own tool loop. It is a thin client in front of a running Philter instance, so you need Philter deployed first.
Pair the MCP server with a locally hosted model. Redaction happens before the model sees the data, but the rest of the conversation still reaches whatever model the client is configured to use.
How they fit together
The products are not alternatives to each other so much as layers, and most of them share the same parts.
LLM Traffic
Drop-in proxy for OpenAI, Anthropic, Amazon Bedrock, and any OpenAI-compatible provider. PII is redacted before prompts leave your network.
Core Redaction Engine
Philter is the self-hosted HTTP API; Phileas is the embeddable library (Java, Python, .NET). Both are powered by PhEye NLP models and driven by Phileas policies.
Discovery at Rest
Crawls local filesystems. Maps where sensitive data lives across your infrastructure before it becomes a compliance finding.
Shared Phileas policy format. One definition applied by all products.
Precision, recall, and F1 measurement. Fail the build when accuracy regresses.
Production PII flow monitoring and anomaly alerting.
Human review, structured exemptions, and audit trail.
Differential privacy for safe aggregate analytics.
For diagrams and deployment topologies, see the architecture overview.
Not sure yet
Run your first redaction with Phileas. It takes about five minutes and the policy format is shared across the toolkit, so nothing you learn is wasted if you move to Philter later.