Aegis

Appearance

Self-hosted · AGPL-3.0Live demo ↗

AEGIS

The self-hosted AI gateway that shows its work

Put guardrails, human approvals and a tamper-evident audit trail between your apps and any LLM — behind an OpenAI-compatible endpoint, configured in one YAML file.

Quickstart →Live demo ↗How it worksWrite a plugin
GuardMask · label · blockPauseA human decidesProveHash-chained evidence
See it work
🔌

Drop-in OpenAI endpoint

Point any OpenAI SDK at /v1/chat/completions. The model field picks an Aegis route; streaming uses standard SSE frames.

REST API

🛡️

Four verdicts, nothing else

Every guardrail returns allow, sanitize, block or require_approval. Every verdict — including allow — is recorded.

Pipeline & verdicts

⏸️

Pause for a human

require_approval checkpoints the run. A named reviewer approves or denies from the CLI, the API or the /approvals page.

Human approvals

🔗

Tamper-evident ledger

Runs and route inventory land in a hash-chained SQLite ledger. Export it and verify the chain offline with aegis audit verify.

Audit & evidence

🧩

Plugin-first, no special cases

PII, residency, classification and budgets are packs built on the same entry-point contracts your own plugins use.

Policy packs

🧪

Scaffold, test, ship

aegis plugin new generates a publishable package with contract tests; aegis plugin test verifies it independently.

Plugin guide

See it work ​

One config, one CLI. A prompt with an API key never reaches a model. A clean question to the loan-underwriting route — whose model runs in a US region — goes straight through; the same route with a Canadian SIN pauses for a named reviewer, who denies it. aegis explain shows why, and the exported ledger verifies offline.

Terminal replay: aegis plugin list shows the policy packs; a prompt with an API key is blocked; a clean underwriting question is answered; one carrying a SIN pauses for approval and reviewer jane denies it; aegis explain shows the verdict trail; aegis audit verify confirms the chain is intact

Try it in your browser: the live demo runs the real guardrails against a mock model — paste a prompt with an email address or a SIN, switch to the underwriting route, and approve or deny the paused run. Nothing to install.

Or run it yourself from a clone with docker compose run --rm dev bash scripts/cli-tour.sh — the config is examples/cli-tour.yaml.

Sixty seconds to a governed endpoint ​

bash
pip install aegis-gateway
aegis init                                 # writes aegis.yaml: fake provider + PII masking
aegis serve --config aegis.yaml --no-auth  # http://localhost:8000
python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dev")
reply = client.chat.completions.create(
    model="default",  # an Aegis route name
    messages=[{"role": "user", "content": "Email jane@example.com the summary"}],
)
print(reply.choices[0].message.content)
bash
curl -s http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"default","messages":[{"role":"user","content":"Hello"}]}'