AI security & governance
Developers sign in once with their work identity. From then on Claude Code, Codex CLI, Claude Desktop and any OpenAI-compatible tool reach Claude and GPT models on Amazon Bedrock through the Aegis gateway — every prompt inspected against a managed content policy and metered, with no API keys to copy, store or rotate, and no prompt content retained.
Prepaid usage. $50 a month per organization. Nothing to install on a server.
Aegis sits between the tools your developers already use and the models they want, and does the governance work that usually gets skipped because it gets in the way.
Developers sign in with your organization's identity provider — Google Workspace today; yours federates in. A small local helper keeps the session fresh in the background. No static API keys exist anywhere. Disable a person upstream and their session expires within its token lifetime; there is no long-lived key to hunt down.
One command, aegis init, points Claude Code, Codex CLI and any
OpenAI-compatible SDK at the gateway, showing every change before it makes it and merging
into the settings you already have. Claude Desktop takes one import inside the app. Nothing
else changes about how anyone works.
Each request carries the developer's identity. A managed content policy covering harmful content and sensitive data inspects prompts before they reach the model and responses on the way back, and usage is metered per developer and per model, to the token — so usage, spend and policy events all have a name on them.
Aegis adds governance to the tools people chose for themselves, rather than replacing them with one they did not.
Claude Code, Codex CLI and OpenAI-compatible clients are configured automatically. Claude Desktop needs one import inside the app, no administrator. Cursor is not supported yet. The ChatGPT desktop app cannot be pointed at Aegis — it signs in with a ChatGPT account rather than as an API client, and no setting changes that.
Prompts are inspected against the Aegis content policy before reaching the model, and blocked when they violate it. Model output is inspected the same way. Inspection is enforced by AWS in a separate account, so it cannot be turned off from the gateway. Prompt and completion content is not retained.
Zero data retention on Tier 1: the AWS account these models run in has its data-retention mode locked to none in every operating region, enforced in an account the gateway cannot edit.
Prompts are inspected against the Aegis content policy before reaching the model, and blocked when they violate it. Inspection is enforced by the gateway rather than by AWS, and model output is not inspected on this path. AWS processes these requests under zero data retention: no request or response data is written to durable storage by AWS or shared with the model provider. We retain no prompt or completion content.
The GPT-5.x models run on a separate AWS surface where the enforcement that makes Tier 1 hold even against a compromised gateway does not exist. Same inspection, weaker enforcement — which is why Tier 2 is off by default and enabled per team by your administrator, with its own description wherever it appears.
The prompts your developers write are your source code, your designs and your plans. Aegis is built so that they pass through, and nothing of them stays behind.
Aegis retains no prompt, completion or tool-output content — not in a log, not in a database, in any environment. What is recorded is metadata: that a policy was violated, which type, by whom and when. Never the content that triggered it.
On Tier 1 the AWS account the models run in has data retention locked to none in every operating region, so Amazon Bedrock keeps nothing either. Tier 2 states its own retention, separately above.
On Tier 1, the content policy is attached to every model call by AWS identity and access management in a separate account that the gateway cannot edit. Even a compromise of the gateway itself could not switch inspection off.
Access is a short-lived token from your identity provider, refreshed by a helper on the developer's machine. There is no static key to leak, commit or forget to rotate. Card details are entered into fields hosted by the payment processor and never reach Aegis.
Governance is only worth having if it covers all of the traffic, all of the time, and cannot be turned off by the thing it governs.
On Tier 1, 100% of prompts pass the managed content policy before they reach the model, and 100% of model responses pass it on the way back. Violations are blocked, not flagged after the fact.
Every request and every tool call carries the developer's identity and is metered per model. Usage, spend and policy events are answerable: who, which model, how much, when.
Policy violations are recorded — what type of violation, by whom, when — never the content that triggered them. Administrators see usage and spend per developer and per model in the portal, and never a transcript.
The GPT-5.x models are inspected by the gateway rather than by AWS, and their output is not inspected. Tier 2 runs under AWS zero data retention, approved for these models on our accounts: no request or response data is written to durable storage by AWS. Keep anything sensitive on Tier 1, and expect Tier 2 to be described separately wherever it appears.
A flat monthly fee for the gateway, and prepaid credit for what your developers actually use.
Covers access to the gateway for your whole team. Your first month's fee is charged one month after you sign up.
Charged when you create your organization and added to your account as usage credit straight away. Top up whenever you like.
Prices in US dollars, exclusive of any applicable taxes. Card payments are taken by our payment processor; card details never reach Aegis.
Search queries are checked against the Aegis content policy by the Aegis gateway before they leave it, and results come from a search index operated by AWS. That check is enforced by the gateway rather than by AWS, so web search carries the GPT-5.x models' weaker guarantee whichever model asked for it.
From nothing to governed traffic in an afternoon. No AWS account of your own, no server to run.
Sign in with Google at the portal, name your organization and enter a card. $50 is charged and added as usage credit, and you become the organization's administrator.
From the portal's Team page, add each developer by their work email. They sign in with the same work identity you did; there is nothing to hand out and nothing to rotate.
On macOS and Windows, developers download the Aegis desktop app from the install page. It walks them through signing in and setting up their tools, and keeps itself up to date.
For Linux, servers and terminal-only use, the same page has a one-line CLI installer for macOS, Linux and Windows. It does the whole job: installs Aegis, signs the developer in, points their tools at the gateway and sends a real request to prove the path works. Most tools then need nothing more than a new terminal.
Honest limits, so the protection you think you have is the protection you have.