Runtime AI Agent Guardrails · LLM Firewall

Deploy AI agents confidently.

Aegis is a runtime firewall for LLM apps and AI agents — guardrails on every prompt, every response, and, where others stop, every tool / MCP call. The control and visibility that open-source alone doesn't provide.

Built on and extending NVIDIA NeMo Guardrails · OpenAI-compatible API

The problem

Agents don't just talk. They act.

The moment an agent can call tools, browse, or reach an MCP server, a bad instruction becomes a real-world action — deleted data, leaked secrets, unauthorized transactions. Prompt-only filters don't see the tool call. That's the gap Aegis closes.

What you get

Control, visibility, and safety over your AI agents.

› Prevent catastrophic actions

Stop agents from misusing tools, deleting data, exfiltrating information, or taking unauthorized actions.

› Block advanced attacks

Defend against prompt injection, indirect injection, goal hijacking, and tool poisoning — especially via MCP.

› Full visibility

See every agent decision, tool call, and policy enforcement in real time.

› Enforce least privilege

Give agents only the exact permissions they need — nothing more.

› Meet compliance

Immutable audit logs and reports mapped to the OWASP LLM Top 10 and the EU AI Act.

› Ship faster, safely

Teams move quickly because the security guardrails are already in place.

How it works

Drop-in in one line. Guardrails on every call.

Point your existing OpenAI-compatible client at Aegis. We run input and output rails, inspect every tool / MCP call against policy, then proxy the clean request to your model — blocking and logging anything dangerous.

1

Repoint your client

Change the base URL and key. No SDK rewrite, no new client — your existing OpenAI code keeps working.
2

Aegis inspects every call

Input & output rails catch prompt injection, jailbreaks, and data exfiltration; every tool and MCP call is checked against least-privilege policy before it runs.
3

Safe traffic reaches your model

Clean requests are proxied upstream to any LLM. Threats are blocked with an OpenAI-compatible response and written to an audit log.

Python · openai SDK

from openai import OpenAI

client = OpenAI(
    base_url="https://aegisapi.graffitecs.com/v1",  # point at Aegis
    api_key="YOUR_AEGIS_KEY",
)

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Summarize this ticket."}],
)

# Blocked requests return an OpenAI-shaped reply with
# finish_reason == "content_filter" — no client changes needed.

Architecture

A firewall in front of your model — and your tools.

Point your existing OpenAI-compatible client at Aegis. Every request passes through the rails; every tool and MCP call is inspected against policy before it runs.

Your app / agent

OpenAI-compatible

Change the base URL + key
No SDK rewrite

Aegis

Rails + tool guard

Input & output rails
Tool / MCP inspection
Least-privilege policy
Audit + analytics

Model provider

Any LLM

Proxied upstream
Streaming supported

Who it's for

Built for teams putting agents into production.

Enterprises and AI-native companies deploying agents where a mistake is expensive — especially in regulated industries.

CISOs & security engineering AI platform teams Finance · Healthcare · Legal · Government AI-native scale-ups MSSPs & security consultancies

Early access

Get Aegis before your agents get loose.

Join the waitlist for early access and design-partner slots. We'll reach out about pilots.