Monday morning, I opened client data in Claude Code
Last Monday at 10 a.m., I was in a shared office in Hangzhou, using Claude Code to review a client interview transcript. The file contained names, email addresses, and pricing. I pasted it in without thinking much. Later, I saw reports from the community saying that when this tool interacts with certain external services, it may also send your email address along with the request.
My first reaction wasn’t to switch tools. I closed my laptop: the data had already gone out, so what now? I had made the “AI is just a chat window” mistake before. I underestimated how account information travels with every request.
What it is, and why it matters
Claude Code is Anthropic’s command-line coding assistant. It can read project files, modify code, and connect to websites or other external services to complete tasks. The issue is that some external sites may receive the email address associated with your login. That doesn’t mean Claude Code deliberately publishes all your client files. It means the email itself may be used to link accounts, identify visitors, or receive service notifications.
My colleague Lin Yue was debugging her quote-generation page with Claude Code on a Friday afternoon in March 2025. She later checked her access logs and found that an external tool had received her work email. Her client data was not exposed because of it, but she changed her email alias that same day and paused external web access.
For those of us without a coding background, the point isn’t to chase the technical terminology. It’s this: anything that hands content to a third party deserves a test run with fake data first.
What it costs to replicate this check today
Money: 0 yuan. Time: about 20 minutes. Technical barrier: enough to log into an account and know your primary email; no coding required. First step: open Claude Code’s settings or permissions page and look for options such as “External Access,” “Web Access,” or “MCP/Tool Connections.” If the labels differ, check the documentation for its third-party service integrations.
Then I do three things: First, I temporarily disable unnecessary internet and external-tool permissions. Second, I separate my work-email alias from my personal email. Third, I create a redacted test folder containing only fictional details such as “Client A, quote: 500 yuan,” with no names, phone numbers, contracts, or real attachments.
Not everyone needs this tool. If you only use the web version for chat and never let it read local projects, I wouldn’t go through the whole exercise. There’s no urgency if you’re not testing it now, but before handling client data, I think those extra 20 minutes are worth it.
Adjust the approach to your stage
If you’re just starting out: I’d use fictional data while I get familiar with it, keep my primary email separate from my work email, and avoid buying any premium service yet.
If you already have one or two clients: I’d pause external web access and run through the workflow once with fake data. For contracts, chat logs, and quotes, I’d return to the files and verify everything manually and locally.
If you’re scaling: I’d have a technical colleague write a one-page checklist of which services are allowed to connect and which files must never be uploaded, then configure permissions separately for each team member. Default settings aren’t proof of safety.
The reminder I took from this was simple: the minutes Claude Code saves me aren’t worth exchanging for my clients’ trust. Which permission would you check first?