PrivacyOctober 4, 2026 · 4 min read

Why we route natively — and what it means for your data

Most multi-AI tools route your messages through their own server, charge you per token, and store everything. Crossagent routes through your browser's own authenticated sessions instead. Here is why that decision changes everything about privacy, cost, and trust.

There are two ways to build a tool that routes messages between AI systems.

The first way: you take the user's message, send it to your own server, your server calls the AI's API with your API key, gets a response, and sends it back. Every message passes through your infrastructure. You need API keys, you bill per token, and you see everything.

The second way: you route through the user's own browser session. The user is already logged into Claude. You inject their message into Claude's input field — the same way they would have typed it. Claude processes it through their existing session. The response comes back through Claude's normal stream. Nothing touches your server.

Crossagent uses the second way. I want to explain exactly why.

The API-proxy model has three problems

Problem 1: It costs you money even when you already paid. You have Claude Pro. You pay $20/month. If I route through my API, I'm calling Claude through a separate API key. Your $20 is useless. I bill you per-token on top. You're paying twice for the same AI.

Problem 2: Your data goes to a server you didn't choose. When you type something sensitive — a business strategy, a personal problem, code from a private repo — and it goes through a proxy, it lives on that proxy's server. In their logs. Under their retention policy. Controlled by whoever operates that server and whoever can compel them legally.

Problem 3: You lose the context of your existing sessions. Claude remembers your previous conversations when you use it directly. When you route through an API, you lose that. You get a blank slate Claude with no memory of your project, your preferences, or your history.

Native routing fixes all three

◯
Your subscription. Your rate limits. Your context.
When Crossagent routes @claude natively, Claude sees it as your message from your account. Your Claude Pro conversation history is intact. Your rate limits apply. Your system prompt, if you've set one, is active. Nothing is bypassed.
◈
Zero bytes of your messages reach our servers.
On the native path, there is no network call to railway.app or any Crossagent endpoint. You can verify this right now: open Chrome DevTools, go to Network, filter by XHR, and type @claude in ChatGPT. Watch the requests. You will see claude.ai. You will not see our server.
⬡
We can't read your messages even if we wanted to.
The architecture makes this technically impossible on the native path. The extension works entirely within your browser's sandbox. Your message goes from your keypress to Claude's input element through Chrome's inter-tab messaging API — a local IPC channel, not a network socket.

What about the memory layer?

Engram — the memory system — does touch our server. When you have a cross-agent conversation, a summary is stored in our Railway backend so it can be recalled in future sessions. This is the only moment data leaves your browser.

A few things to know about how Engram handles this:

  • Your memories are tied to a random client-generated ID with a ca_ prefix. No email, no name, no account required.
  • You can delete any individual memory from the panel with a single click.
  • You can wipe everything at once via the API: DELETE /memory/delete/all?userId=
  • The backend code is MIT licensed and self-hostable. If you want your memories on your own infrastructure, you can point the extension at your own deployment.
  • Native routing — the @mention calls — never touches Engram. Those are completely separate systems.

You can use Crossagent with Engram disabled entirely and still get unlimited @mention routing. The memory layer is opt-in by design.

The test

Here is how I verify that native routing is real every time I make a change:

Disconnect from the internet. Open Claude and ChatGPT in separate tabs. They've already loaded. Type @claude test in ChatGPT.

It works. Claude responds. Because the routing was always happening inside your browser, not through a server somewhere.

That's the whole point. Your data is yours. Your subscriptions are yours. Crossagent just connects them.

Written by

Tanmay Sangam

Builder, Crossagent  ·  Northeastern University  ·  Boston

No API key. No proxy. No extra cost.

Install Crossagent and use your existing AI subscriptions the way they were meant to work.

Add to Chrome — free →