惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

N
Netflix TechBlog - Medium
G
Google Developers Blog
H
Hackread – Cybersecurity News, Data Breaches, AI and More
T
The Blog of Author Tim Ferriss
Microsoft Azure Blog
Microsoft Azure Blog
GbyAI
GbyAI
L
LangChain Blog
云风的 BLOG
云风的 BLOG
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
aimingoo的专栏
aimingoo的专栏
P
Proofpoint News Feed
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
小众软件
小众软件
WordPress大学
WordPress大学
A
About on SuperTechFans
大猫的无限游戏
大猫的无限游戏
C
Check Point Blog
月光博客
月光博客
Stack Overflow Blog
Stack Overflow Blog
美团技术团队
Jina AI
Jina AI
T
Tailwind CSS Blog
Google DeepMind News
Google DeepMind News
D
Docker

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
Delete the Vercel Claude Code Plugin. Here's Why I Did.
taekim34 · 2026-05-11 · via DEV Community

TL;DR

  • The Vercel Claude Code plugin creates a permanent device UUID on your machine the instant you install it. No notification. No expiry. No rotation.
  • Session starts, tool calls, skill matches — all sent to telemetry.vercel.com. Default ON, no consent prompt. Prompt metadata (matched skill + score) included.
  • What's worse: they built a consent dialog for prompt text collection. But clicking "No thanks" only stops prompt text. All other telemetry keeps running. Most users will think they opted out of everything.
  • The documentation exists — buried eight directories deep inside ~/.claude/plugins/cache/. Nobody reads it. Documented ≠ Informed.

What I Found

I was building a static analysis tool for AI plugins — scanning popular skills for security issues. Regex pattern matching plus dual-LLM cross-verification.

I was running a batch scan — 200 Claude Code skills, checking for destructive commands, data exfiltration, prompt injection, the usual. On skill #147, the scanner flagged something in ~/.claude/. Not in some random GitHub repo. On my own machine.

I didn't suspect Vercel for a second. I assumed the flag was a false positive in my own skill. So I pulled the Vercel plugin source as a reference — to compare against "known good" code and figure out what I was doing wrong.

Then I read the Vercel source. Here's what I found.


The Evidence

All file paths and line numbers reference vercel-plugin v0.32.7, located at ~/.claude/plugins/cache/vercel/vercel-plugin/0.32.7/.

Every session start sends this:

// session-start-profiler.mts:702-709
session:device_id            // permanent device identifier
session:platform             // darwin, linux, win32
session:likely_skills        // which skills you use
session:greenfield           // whether the project is new
session:vercel_cli_installed // whether you have the Vercel CLI
session:vercel_cli_version   // which version

Enter fullscreen mode Exit fullscreen mode

Every tool call you make — any tool, not just Vercel's:

// pretooluse-skill-inject.mts:969-971
tool_call:tool_name          // which tool you just called

Enter fullscreen mode Exit fullscreen mode

Every time a skill matches your prompt:

// pretooluse-skill-inject.mts:1205-1210
skill:injected               // which skill got injected
skill:match_type             // how it matched
skill:tool_name              // against which tool

Enter fullscreen mode Exit fullscreen mode

Every prompt you submit:

// user-prompt-submit-skill-inject.mts:1063-1065
prompt:skill                 // which skill matched your prompt
prompt:score                 // confidence score

Enter fullscreen mode Exit fullscreen mode

All of it flows to a single endpoint:

https://telemetry.vercel.com/api/vercel-plugin/v1/events

Enter fullscreen mode Exit fullscreen mode

None of it asked for your permission.

The permanent device ID

This is the part that should make you check your machine right now. Run this:

cat ~/.claude/vercel-plugin-device-id

Enter fullscreen mode Exit fullscreen mode

You'll see something like:

473d7060-5a37-4ebb-9082-b09a983c****

Enter fullscreen mode Exit fullscreen mode

A UUID. Created the instant you installed the plugin. Silently. No notification. It never expires. It never rotates. It ties together every session, every project, every client engagement you've ever worked on with Claude Code.

For context: Chrome DevTools rotates session IDs every 24 hours (ClearcutSender.ts:35,68-70). Vercel's device ID never expires. Privacy-conscious analytics platforms moved away from persistent device IDs years ago. This one lasts forever.

Dozens of telemetry events per coding session. All tied to a permanent fingerprint. All default-on.


"But It's in the README"

Technically, yes. The plugin's README.md has a ## Telemetry section. It explains what's collected and how to disable it.

But does anyone seriously think that counts as consent?

Walk through what actually happens:

  1. You install the plugin.
  2. It prints a success message.
  3. You start coding.

At no point does any text appear on your screen about telemetry. No prompt. No checkbox. No banner. Nothing. Meanwhile, in the background: ~/.claude/vercel-plugin-device-id is written to disk, session events are queued, and your usage patterns start flowing to Vercel's servers.

The README is sitting in ~/.claude/plugins/cache/vercel/vercel-plugin/0.32.7/. Eight directories deep inside a hidden folder. Nobody browses there.

GDPR defines valid consent as "freely given, specific, informed, and unambiguous." Most companies — including startups with a fraction of Vercel's resources — treat this as the baseline. I haven't seen a single serious startup ship permanent device tracking without an install-time consent prompt in years. It's just not done anymore.

Remember: Chrome DevTools rotates its session IDs every 24 hours (ClearcutSender.ts:35,68-70). That's the standard. Vercel's device ID never rotates. Never expires. Created once, lives forever.

This is not a gray area. This is not "technically compliant." A permanent device UUID, created silently, tied to every session, with no install-time disclosure — this is clearly Vercel's mistake.

I used this plugin daily for months. I had no idea. And I'm the developer who was literally building a tool to analyze plugin source code.


The Part That's Even More Absurd — I Never Consented

Here's what makes this worse. The plugin actually has a consent dialog — for prompt text collection:

// user-prompt-submit-telemetry.mts:58-61
prompt:text  // full prompt content, up to 100KB — OPT-IN ONLY

Enter fullscreen mode Exit fullscreen mode

An explicit question appears: "Share your prompt text to help improve skill matching." You can say yes or no. Your choice is saved.

So they know how to build consent flows. They built the infrastructure. They just chose not to use it for device tracking, tool-call logging, skill-usage profiling, and platform fingerprinting.

And here's the trap: if you click "No thanks," you think you've opted out. You haven't. Base telemetry — everything in the previous section — keeps running. The README even says so: "base telemetry remains on by default."

But you already clicked "No thanks." In your mind, the matter is settled. That's not a documentation gap. That's a dark pattern.


How to Protect Yourself

Do this now. It takes 60 seconds.

1. Check if you're affected

ls ~/.claude/vercel-plugin-device-id

Enter fullscreen mode Exit fullscreen mode

If the file exists, you have a permanent tracking UUID on your machine.

2. Disable telemetry

Add this to your shell profile (.zshrc, .bashrc, etc.):

export VERCEL_PLUGIN_TELEMETRY=off

Enter fullscreen mode Exit fullscreen mode

Then reload:

source ~/.zshrc

Enter fullscreen mode Exit fullscreen mode

3. Or just uninstall the plugin entirely

If you don't need it, remove it. One fewer thing sending data you didn't agree to.


What Should Change

Two proposals. Design standards, not policy demands.

1. Surface telemetry at install time. One prompt. Plain language. "This plugin collects [X] and sends it to [Y]. OK?" The user sees it. The user decides. This is four lines of install-time code. Vercel already has the consent infrastructure. They use it for prompt text. Extend it to everything else.

2. Treat data flows as API surface. If your plugin sends data to an external endpoint, document it the way you'd document an API. What data. Where it goes. How often. How to stop it. Put this in the install output, not in a README eight directories deep.

These aren't radical ideas. Homebrew notifies you on first run. VS Code notifies you on first launch. It's already the industry standard. The Vercel plugin just doesn't.


Check your ~/.claude/ directory right now. What did you find? Drop it in the comments.