Anthropic's IDE is shipping with plugins that want to read every word you type, and developers are bailing for alternatives.
The Summary
- Vercel's plugin for Claude Code sends your prompts and file paths to their telemetry servers, raising privacy concerns for developers using AI coding assistants
- One developer publicly moved their $100/month Claude spend to Zed + OpenRouter over telemetry and locked-in pricing
- The backlash reveals a trust problem: developers building the agent economy don't trust the tools building it
The Signal
Akshay Chugh dug into the Vercel plugin code and found it's collecting prompts, file paths, and project context, then shipping it off to Vercel's servers. This isn't some edge case buried in the ToS. The plugin explicitly requests permission to read your prompts. The problem is most developers aren't reading those permission screens, they're just clicking through to get the autocomplete working.
This matters because Claude Code isn't some weekend hobby project. It's positioning itself as the professional IDE for the agent era. Developers are paying $100/month for it. When that developer switched to Zed with OpenRouter, they got the same model access with more control and transparency, no lock-in to Anthropic's billing. The telemetry issue was the trigger, but the locked pricing was the underlying problem. You're renting access to models you could access elsewhere, and the plugins you install are mining your workflow.
The pattern here is familiar from Web2: free tools that feel magical at first, then quietly start extracting value from your data. Except now the data isn't your photos or your social graph. It's your actual work product, your prompts, the proprietary code you're shipping. The developer community's reaction on Hacker News was sharp, with 264 upvotes and developers comparing notes on alternatives. This isn't just one person being paranoid.
The deeper issue is trust architecture. In Web3, you can verify what a smart contract does because the code is on-chain. In Web4, you're supposed to trust that the AI coding assistant's plugins aren't shipping your IP to third parties. That's a step backward. Developers building agent systems want agents they control, not agents that report back to corporate servers.
The Implication
If you're using Claude Code or any AI IDE, audit what plugins you've installed and what permissions you've granted. Check if there's a local-only or open-source alternative that gives you the same capabilities without the telemetry tax. The Zed + OpenRouter combo is one option, but the broader lesson is to treat these tools like you'd treat any SaaS platform: assume the default is extraction, verify the permissions, and be ready to move if the value equation shifts.
For companies building developer tools in the agent era, this is a warning shot. Developers will tolerate some data collection if the value is clear and the controls are legit. But if it feels like surveillance dressed up as helpful analytics, they'll route around you. The companies that win here will be the ones that let developers own their toolchain, not rent it.