MCP server not working in Cursor 1.0

Bug Reports
1

Describe the Bug

Hi I just updated to the new version this morning and the playwright MCP server that was working yesterday is not working today. It is not opening the browser and I can’t follow the agent actions and the agent seems to be getting partial output because it not able to finish actions that yesterday could do. It may be a problem with MCP servers in general and not this one in particular, but this is my config for it:

    "playwright": {
      "command": "pnpm",
      "args": ["dlx", "@playwright/[email protected]"]
    },

Steps to Reproduce

  1. Add playwright to MCP servers:
    "playwright": {
      "command": "pnpm",
      "args": ["dlx", "@playwright/[email protected]"]
    },
  1. Some tools not visible in Cursor settings
  2. Asking the model to do simple tasks in the browser → clicks don’t work, browser window never opens o the model get errors back
  3. The model get confused while trying to accomplish the task, even hallucinates buttons or thinks that is in a different page.

Expected Behavior

I was able to try again the same day with version 0.5 and the mcp tools load and the model is able to click and navigate and accomplish tasks relaunching the exactly same prompt and configuration.

Operating System

Linux

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.0.0
VSCode Version: 1.96.2
Commit: 53b99ce608cba35127ae3a050c1738a959750860
Date: 2025-06-04T19:26:40.367Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Linux x64 6.8.0-60-generic

Does this stop you from using Cursor

No - Cursor works, but with this issue

2

Hey, it works well for me with both npx and pnpm, so it might be an issue with the package manager.

3

I am having the same issue with puppeteer

4

The example is using pnpm dlx, which is more or less equivalent to npx, I tried with npx too, It looks like it works, but the model answers and actions are bad or never get anywhere, the first impression is that the model is not understanding the task, but when you try with the same model, same prompt and mcp config in cursor 0.5 it works. So I think it is a problem in the MCP integration or how the tools or the output are provided to the model.