Where does the bug appear (feature/product)?
Cursor CLI
Describe the Bug
On a webapp test project i want to test the ability of the CLI agent to run e2e tests from a simple prompt.
It works from the integrated cursor chat (ctrl+L) but not with the agent CLI tool.
It seems there is something wrong related to X server and ubuntu in headless mode. It is maybe not something involving directly cursor and agent, maybe it is a missing dependency on my system but i can’t get it. Or maybe i need to open this issue to the chrome-devtools-mcp repo but i start here as my entry point is cursor.
Here is the kind of output i get from my agent command (see steps to reproduce for details) :
{"type":"result","subtype":"success","is_error":false,"duration_ms":140272,"duration_api_ms":140272,"result":"\nChecking open pages, then navigating to http://localhost:3000.\n\nNavigating to http://localhost:3000:\n\nChecking the project structure and whether the app is running:\n\nServer is running. Checking Chrome DevTools MCP configuration for headless mode:\n\nChrome must be running in debug mode. Checking if it's already running, then starting it if needed:\n\nChrome is running in debug mode. Listing pages, then navigating to the app:\n\nChecking the login page structure, then retrying the MCP tools:\n\nTrying to work around the X server issue. Checking if we can use xvfb-run or configure the environment:\n\nDISPLAY is set, but the MCP server still reports an X server issue. Trying to use the Chrome DevTools Protocol directly to interact with the page. First, checking available pages via the debugging port:\n\nRetrying the MCP tools with a timeout. If that fails, we'll use a direct Chrome DevTools Protocol script:\n\nCreating a script using Puppeteer to interact with the page, since the MCP server has an X server issue:\n\nCreating a Puppeteer script to automate the login and verify the link-list page. Checking if Puppeteer is available:\n\nInstalling Puppeteer and creating a script that connects to the existing Chrome instance:\n\nCreating a Deno script using the Chrome DevTools Protocol to automate the login:\n\nTrying a simpler approach: using the Chrome DevTools Protocol via HTTP to navigate and execute JavaScript. Creating a more practical script:\n\nRewriting the script to use WebSocket properly. Checking for a Deno-compatible WebSocket library:\n\nChecking the Chrome DevTools Protocol endpoint response and fixing the script:\n\nUpdating the script to use PUT for creating pages and to work with existing pages:\n\nRewriting the script to use WebSocket with Deno's native WebSocket API:\n\nRunning the script to test the login:\n\nThe form isn't found. Checking what's on the page and adjusting the wait:\n\nRerunning the script to inspect the page content:\n\nThe result is undefined. Fixing error handling and response parsing:\n\nRerunning to inspect the response structure:\n\nThe form is found. Fixing the script to access the nested result structure correctly:\n\nRunning the updated script:\n\nTask completed. The script:\n1. Connected to Chrome via the DevTools Protocol\n2. Found the login form on the page\n3. Filled in email: `[email protected]` and password: `1234`\n4. Submitted the form\n5. Verified navigation to the link-list page\n6. Confirmed the link-list page is loaded\n\nCleaning up the temporary test script:\n\nTask completed:\n\n1. **Loaded the app** on http://localhost:3000\n2. **Logged in** with `[email protected]` and password `1234`\n3. **Verified** the link-list page is loaded\n\nThe script confirmed:\n- Login form found and filled\n- Form submitted successfully\n- Navigation to `#/link-list` completed\n- Link-list page elements are present\n\nNote: The Chrome DevTools MCP server had an X server configuration issue, so I used a Deno script with the Chrome DevTools Protocol to complete the task. The result is the same—login successful and the link-list page is loaded.","session_id":"3c02242f-40c5-4d49-ac67-82c9a6b37617","request_id":"5c2b202d-620f-4bde-8b1e-99382381f8b0"}
So when reading carefully, we see that the MCP server reports an X server issue. Then finally the agent bypass the issue by writing a test script and the task is done but not in the way i want to do it at all.
Steps to Reproduce
My Linux is Ubuntu 24.04.3 LTS.
I use chrome-devtools-mcp.
Notice my js/ts platform is deno 2.3.7 to build and serve my webapp which is a very simple vanilla TS webapp without any dependency instead of vite and tailwind.
When using the chat (ctrl+L) inside cursor, everything works fine.
The chrome-devtools-mcp is properly configured under ./cursor/mcp.json.
The chrome browser is running properly with google-chrome --remote-debugging-port=9222 --user-data-dir=/tmp/chrome-debug
I give this kind of prompt to the chat :
- load the app on http://localhost:3000
- login with [email protected] and 1234
- on the list page add the following link :
- name = Google
- url = https://www.google.fr
- no description
- go the the search page and search for any link with the go string
- check that the link Google that you added before exists in the result
- makes another search with the za string
- check that the result list is empty
And i can see the browser running and controlled by the agent. Then the agent write a resume of the test. It’s OK for that mode.
But with agent CLI tool it does not work.
To run the agent i use this script :
#!/bin/bash
source .env
echo "starting test"
agent --api-key $CURSOR_API_KEY --model "composer-1" -p --force --output-format json "load the app on http://localhost:3000 then login with [email protected] and 1234 then check that the link-list page is loaded."
The prompt is simpler for testing purposes. After a minute or two i get the output given in the describe the bug section.
I have also tested with agent in interactive mode, i have exactly the same issue as with the non-interactive mode using -p option with agent.
Expected Behavior
I don’t want the agent to write a puppeteer script to bypass the X server issue, i want my chrome to run in headless mode properly.
Operating System
Linux
Current Cursor Version (Menu → About Cursor → Copy)
Version: 2.3.21
VSCode Version: 1.105.1
Commit: 68e0a0385b87408d050869ea543e3778ad53f780
Date: 2026-01-02T23:46:13.381Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Linux x64 6.14.0-37-generic
For AI issues: which model did you use?
Composer-1
Does this stop you from using Cursor
No - Cursor works, but with this issue
