Aegis

Private AI in one command. Strip PII from every LLM prompt, locally.

Try it

docker run -p 8080:8080 ghcr.io/svenplb/aegis

Use it

export OPENAI_BASE_URL=http://localhost:8080/v1

That's it. Your prompts are now private.

What it does

Aegis sits between your app and any OpenAI-compatible API. It strips personal data from prompts before they reach the LLM, and restores it in responses. Everything runs locally. Your data never leaves your machine.

Your App ──→ Aegis Proxy ──→ LLM API
               │
          strips PII from prompts
          restores PII in responses

Before (what hits the LLM):

Send the contract to alice@example.com.
Her SSN is 123-45-6789 and she lives
at Hauptstraße 42, 10115 Berlin.

After (what Aegis sends instead):

Send the contract to [EMAIL_1].
Her SSN is [SSN_1] and she lives
at [ADDRESS_1].

The LLM never sees the real data. Aegis restores it in the response before returning to your app.

Works with

• OpenAI API
• Ollama
• LM Studio
• llama.cpp
• Any OpenAI-compatible API

Configuration

All via environment variables. No config files.

Variable	Default	Description
AEGIS_UPSTREAM_URL	https://api.openai.com	LLM API to proxy to
AEGIS_PORT	8080	Listen port
AEGIS_LOG_LEVEL	info	debug / info / warn / error
AEGIS_CACHE_TTL	1h	Conversation mapping TTL
AEGIS_CACHE_SIZE	1000	Max cached conversations
AEGIS_UPSTREAM_TIMEOUT	60s	Upstream request timeout
AEGIS_MAX_BODY_SIZE	10485760	Max request body (bytes)
AEGIS_SCANNER_URL	(disabled)	Ollama URL for LLM-based scanning
AEGIS_SCANNER_MODEL	gemma3:4b	LLM model for sensitivity detection
AEGIS_SCANNER_TIMEOUT	30s	LLM scanner request timeout

Example: Ollama

docker run -p 8080:8080 -e AEGIS_UPSTREAM_URL=http://host.docker.internal:11434 ghcr.io/svenplb/aegis

LLM Scanner (opt-in)

For deeper detection, point Aegis at a local LLM via Ollama. The LLM catches contextual sensitivity that regex can't: business secrets, emotional content, relationship context, medical implications.

AEGIS_SCANNER_URL=http://localhost:11434 AEGIS_UPSTREAM_URL=https://api.openai.com aegis

The regex scanner runs first (<5ms), then the LLM scanner fills in the gaps (~6-16s depending on model). Best-effort: if the LLM is slow or down, requests proceed with regex-only results.

What it detects

16 PII types with format-specific validation:

Type	Examples	Validation
EMAIL	alice@example.com	RFC 5322
PHONE	+49 170 1234567	International + local
SSN	123-45-6789	20+ national formats
IBAN	DE89 3704 0044 0532 0130 00	MOD-97 checksum
CREDIT_CARD	4111-1111-1111-1111	Luhn algorithm
IP_ADDRESS	192.168.1.1	IPv4 + IPv6
URL	https://example.com	Pattern match
DATE	15.03.1990	Multiple formats
ADDRESS	Hauptstraße 42, Berlin	Context-validated
FINANCIAL	EUR 1.500,00	15+ currency formats
And more...	SECRET, MEDICAL, AGE, ID_NUMBER, MAC_ADDRESS, GPS

Strong EU coverage: Germany, Austria, Switzerland, France, Italy, Spain, Netherlands, Belgium, Poland, Portugal, Sweden, Denmark, Finland, Norway, Czech Republic, Slovakia, Ireland, UK.

How it works

1. Your app sends a request to Aegis (same format as OpenAI API)
2. Aegis scans every message for PII using regex patterns with format-specific validators
3. PII is replaced with deterministic tokens ([EMAIL_1], [PERSON_1], etc.)
4. The clean request is forwarded to the actual LLM
5. The LLM response comes back with tokens
6. Aegis restores the original PII in the response
7. Your app gets the response as if it talked to the LLM directly

System prompts are never scanned (they contain your instructions, not user data).

Privacy guarantees

• Zero data exfiltration. The proxy makes no outbound calls except to your configured upstream.
• No logging of PII. Even at debug level, message content is never logged.
• No telemetry. No analytics, no tracking, no phone-home.
• No disk writes. All state is in-memory. Restart = clean slate.
• Fail closed. If PII scanning fails, the request is blocked. Unscanned data is never forwarded.

Build from source

go install github.com/svenplb/liebotsch/cmd/aegis@latest

Or:

git clone https://github.com/svenplb/liebotsch.git
cd liebotsch
go build -o aegis ./cmd/aegis/
./aegis

License

MIT