The Middleman Attack: Why Your AI Wrapper is Logging Your Code

If you aren't using your own API key, you are trusting a middleman.

That free "AI Wrapper" app you downloaded sits right between you and the provider. It can read your messages, log your code snippets, and see your API keys. And you agreed to let it.

There is a massive privacy firewall between consumer apps (ChatGPT, Claude.ai, Gemini) and their developer API equivalents. Most people don't realise they're agreeing to two completely different Terms of Service depending on how they access the same models.

---

The Breakdown

OpenAI (ChatGPT vs. API)

• ChatGPT (Consumer): By default, OpenAI uses your conversations, code snippets, and uploads to train future models. You can opt out in settings, but the default is opted in.
• API (Developer/BYOK): Data submitted through the API is not used to train models. Retained for 30 days for abuse monitoring only. Zero-day retention is available for sensitive use cases.

Anthropic (Claude vs. API)

• Claude (Consumer): As of late 2024, consumer chats can be used to improve future models unless you manually opt out. If you leave it on, they retain data for up to 5 years.
• API (Commercial): Not used for training. Your data belongs to you.

Google (Gemini App vs. API)

• Gemini App (Consumer): Chats may be reviewed by human reviewers and used to enhance Google products.
• Gemini API (Paid): Data is not used to train foundational models. Logged temporarily for policy violation checks only.

---

Why This Matters for Developer Tools

When you use a standard AI wrapper app where they provide the AI, they're often operating under the consumer terms — or worse, they're logging your data on their own servers before forwarding it to the API.

Your codebase, your API keys, your internal documentation — all passing through a middleman.

How AICoven Handles This

AICoven is a Bring Your Own Key (BYOK) client. Your traffic does go through our backend — we need it to hold thread history, run tools, stream turns, and build agent context. That's not the same as a wrapper that stores your conversations in plaintext and reads them for analytics or training.

What we're trying not to be:

• A logging middleman that keeps a browsable copy of your code and sells "AI features" on consumer ToS.
• A key custodian that routes your prompts with their API key so they inherit the wrong privacy tier.

What BYOK gives you:

• Developer-grade provider terms by default. Your API key means the model vendor sees you under the developer/API policy, not ChatGPT consumer defaults.
• Your keys, encrypted at rest in our vault — not sitting in a plaintext column waiting for a curious employee.

For message content on the cloud path: we encrypt with your encryption key. We don't keep a plaintext archive of your chats for humans to read. Content is decrypted in RAM only long enough to orchestrate the turn (context assembly, tool routing, forwarding to the provider you chose), then handled under our retention model — not mined for product analytics.

That is different from "your packets never touch our servers." They do. Hermes routing, Mac shell tools, and streaming all run through the same orchestration layer — see how we route cloud, Hermes, and your Mac. The claim is narrower: we're not the kind of middleman that treats your thread as their training data.

If you care about what happens to your code after you hit Enter, check how your current tool handles it — whether you're on the consumer tier or the API tier, and whether the app stores readable copies of everything you typed.