Privacy mode (local)

For workloads where data residency or privacy compliance rules out a hosted gateway. Everything runs in a Docker container on your own hardware. The only outbound traffic is the LLM call itself, to the provider you choose.

This recipe assumes local mode. Cloud mode by definition involves a hosted service.

What this pipeline gives you

All caching, learning, and context management happens in a SQLite file on your disk.
(v1.1) airgap enforces that no other network call can be made by the gateway.
(v1.1) guardrails redacts PII before the request leaves your machine for the provider.

The pipeline

Env var


docker run -d \
  -p 3099:3099 \
  -v ~/.prxy:/data \
  -e ANTHROPIC_API_KEY=sk-ant-xxx \
  -e PRXY_PIPE='ipc,patterns,semantic-cache' \
  prxymonster/local:latest

When v1.1 ships:


PRXY_PIPE='airgap,guardrails,ipc,patterns,semantic-cache'

Why this order

(v1.1) airgap first — enforces no-network-out at the start of every request. Other modules then run inside the network sandbox.
(v1.1) guardrails — redacts PII before anything else touches the prompt.
ipc — manages context length using only local storage.
patterns — your forged patterns stay in ~/.prxy/prxy.db. Never sync, never share.
semantic-cache — embeddings + cached responses live locally.

What’s stored on disk


~/.prxy/
├── prxy.db          ← SQLite. Patterns, cached embeddings + responses, sessions.
├── blob/            ← Archived (compressed-out) message bodies.
└── config.yaml      ← Optional pipeline config.

The container has no other state. Restart it — everything persists. Delete ~/.prxy/ — everything is gone.

What leaves your machine

The LLM API call itself. Nothing else.

The gateway makes outbound calls only to:

The configured provider (Anthropic, OpenAI, Google, Groq) — over HTTPS.
Optionally: an embedding provider (Voyage / OpenAI) for semantic-cache’s vector lookups. Disable by leaving VOYAGE_API_KEY and OPENAI_API_KEY unset for embeddings — mcp-optimizer and semantic-cache fall back to the offline stub embedder.

No telemetry. No phone-home. No background sync. The container’s network egress can be audited with any standard packet capture tool.

Hardening checklist

No OPENAI_API_KEY set unless you actually want OpenAI to receive your prompts.
VOYAGE_API_KEY unset → falls back to offline stub embeddings.
(v1.1) airgap module added → enforces zero non-provider network traffic at the gateway layer.
Bind the container to localhost only: -p 127.0.0.1:3099:3099.
Ship the ~/.prxy/ directory to encrypted backups; nowhere else.
Audit outbound traffic via tcpdump or your firewall after first-run sanity test.