Security and data handling

Where your data goes. Who holds the key. What gets logged.

Three questions worth answering before you pay anyone to put an AI assistant on your website. Here are our answers, written down, in enough detail that you can check them yourself.

Last reviewed 10 July 2026 · Applies to SiteAgent builds

The short version

If you read nothing else on this page.

Where does my data go?

To Anthropic, and nowhere else. A visitor's message travels from their browser to a proxy running on your infrastructure, and from that proxy to Anthropic's API. It does not pass through a DataBridge Ops server. We do not operate a service in the middle of your traffic. After delivery, there is no path from your deployment back to us.

Who holds the API key?

You do. The key is issued on your own Anthropic account, stored as a secret in your own infrastructure, and used only by your own proxy. It is never written into any file a website visitor can read. We do not retain a copy after handoff.

What does the proxy log?

Nothing, by default. The proxy we ship does not write conversation content to storage. It holds a message in memory long enough to forward it and return the reply. If you want logging, we can add it, and we will tell you exactly what it captures and where it lands before we build it.

Architecture

One hop through a proxy you own.

Most AI widget implementations put the API key directly in the browser, inside the JavaScript the page loads. Anyone with developer tools open can read it, copy it, and spend against your account. That is not a configuration mistake. It is the shape of the design.

SiteAgent uses a different shape. The browser holds no credentials at all. It sends the visitor's message to a small server-side proxy on your infrastructure. The proxy attaches the key, enforces the limits, and forwards the request to Anthropic. The reply travels back the same way.

The proxy runs as a Cloudflare Worker, a serverless function, or on your own backend. Whichever you already run.

Server-side proxy

The key stays server-side

Every request from the visitor's browser goes to a proxy on your infrastructure. The proxy adds the API key before forwarding to Anthropic. The key lives in a secret store, not in any file a visitor can inspect.

Origin locking

Only your domain can call it

The proxy checks the request origin against an allowlist of your domains. A request from any other origin is rejected before it reaches Anthropic. Somebody who finds your endpoint URL still cannot use it.

Server-side token cap

The cost ceiling is enforced at the proxy

The maximum tokens per response is set by the proxy, not by the browser. Even if someone edits the request in transit, the proxy overrides it. No browser-side manipulation can run up your bill.

Stateless sessions

Nothing persists between conversations

Each conversation is isolated and held in the browser tab only. Nothing accumulates, nothing is written to a database. There is no store of visitor conversations to breach, because no such store is created.

Key custody

Your account, your key, your bill.

SiteAgent does not resell API access. You create an Anthropic account, you generate the key, you set the spend limits, and Anthropic bills you directly. This matters for a reason that has nothing to do with cost: it means the credential that controls your spend was never in anyone else's hands to begin with.

→You create the key. On your own Anthropic account, at build time. We tell you exactly which permissions it needs.

→You install the key. As a secret in your own infrastructure. If you would rather we do it, you grant temporary access and revoke it at handoff. If you would rather do it yourself, we send instructions and you paste it in. Both are normal.

→You can rotate it at any time, without us. Replace the secret, and the deployment keeps working. Nothing in the widget or the proxy is bound to a specific key.

→We keep no copy after handoff. If we held the key during the build, we tell you when it can be rotated, and we recommend you rotate it.

→You set the spend cap in the Anthropic console. The proxy's token ceiling is a second limit underneath it, not a substitute for it.

Visitor data

What is captured, what is not, and for how long.

A visitor types a question. That text has to reach a language model to be answered. That is the whole of what happens, and it is worth being precise about the parts.

Item	What happens to it
Visitor's message text	Sent from the browser to your proxy, forwarded to Anthropic, answered. Held in the browser tab for the length of the conversation. Not written to any database by the code we ship.
Conversation history	Exists only in the open browser tab. Closing or refreshing the page ends it. The next visit starts from nothing.
Names, emails, or anything else a visitor types	Treated exactly like any other message text. The assistant is instructed not to solicit personal information and to route people to your intake form instead. It is a routing layer, not a collection point.
Your system prompt	Lives at the proxy, on your infrastructure. It is a plain text file describing your business. It is never sent to the browser, so a visitor cannot read your instructions.
IP addresses and request metadata	Whatever your own hosting platform records as part of serving a request, under its own policy. We add no separate collection.
Anything DataBridge Ops receives	Nothing, after delivery. We do not run a server in your request path, and we do not receive telemetry from your deployment.

Anthropic processes the message in order to generate a reply, under its own commercial terms and data policy. That relationship is between you and Anthropic, on your account. If your industry requires a specific data processing agreement, you sign it with them directly, which is the point of not putting a middleman in the path.

If you want conversation logging, for lead capture or for quality review, we build it. We will tell you what it captures, where it is stored, who can read it, and what the retention period is, before you approve the build. Logging is a decision you make on purpose, not a default you inherit.

Verify it yourself

Do not take our word for any of this.

This website runs SiteAgent. The widget in the corner of this page is the product being described. Every claim above is checkable from your own browser, right now, in about a minute. If you are not technical, forward these steps to whoever is.

One minute, your browser, our code

Look for the API key. You will not find one.

If a widget on any website exposes its key, this is where it shows up.

Open developer tools. F12 on Windows, or Cmd + Option + I on a Mac. Select the Network tab.
Open the chat widget on this page and send it a message.
Find the request that appears in the network list. Open it and read the request headers. There is no x-api-key, and no authorization header carrying a credential. The browser has nothing to send.
Now switch to the Sources tab and open agent.js. That is the entire widget, unminified. It contains one endpoint URL and no secrets.
For contrast, run the same four steps on any AI chat widget you are currently considering. The comparison is the argument.

This is the same architecture every SiteAgent build ships with. It is not a diagram of an intention. You are reading this page through a running deployment of it.

Honest limits

What SiteAgent is not.

A security page that only lists strengths is a marketing page. Here is the other half, so that nothing on this list arrives as a surprise later.

✕Not a certified or audited product. DataBridge Ops holds no SOC 2 report and no ISO certification. If your procurement process requires one from every vendor, we will not clear it, and you should know that before you pay rather than after.

✕Not a system of record. The assistant answers questions and routes visitors. It does not hold customer data, and it should not be given any.

✕Not a guarantee against a wrong answer. The assistant is scoped to what you tell it and instructed to decline what it does not know. It is a language model. It can still be wrong. Do not use it as the only source for anything that carries legal, medical, or financial weight.

✕Not a shield against your own infrastructure. The proxy is as secure as the platform it runs on and the secrets hygiene of the people with access to it. We harden what we ship. We cannot harden your account.

✕Not immune to abuse, only expensive to abuse. Origin locking, message caps, input truncation, and a server-side token ceiling make the endpoint a poor target. They do not make it an impossible one. Keep your spend cap set in the Anthropic console.

If something goes wrong

Who you email, and how fast we answer.

Write to build@databridgeops.com. Every email is answered within one business day, and a suspected key exposure or a live security issue in something we built is answered the same day. There is no ticket system to route around, and no account manager between you and the person who wrote the code.

If you believe a key has been exposed, rotate it first in the Anthropic console, then email us. Rotation takes precedence over diagnosis. Nothing in your deployment breaks when you replace the secret.

If you have found a vulnerability in code we shipped, tell us and we will fix it and tell every affected client. We would rather hear it from you than from a bill.

Now go read the price.

One fixed-price build. You own the code and the proxy at the end of it.

See SiteAgent