MCP client for `playwright` failed to start: program not found

Same thing happened to me. All MCP servers seem to be bugged or something. Here's what I added to my config.toml

[mcp_servers.supabase]
command = "npx"
args = ["-y", "@supabase/mcp-server-supabase@latest", "--access-token", "my_access_token"]

I even tried doing this:

[mcp_servers.supabase]
command = "cmd"
args = ["/c", "npx", "-y", "@supabase/mcp-server-supabase@latest", "--access-token", "my_access_token"]

But got:

MCP client for supabase failed to start: program not found

@SphaeroX
First, run Playwright in the terminal—you can stop it afterwards, and then it works. There might be an issue that Playwright starts up more slowly than Codex, so they’re not synchronized

@SphaeroX First, run Playwright in the terminal—you can stop it afterwards, and then it works. There might be an issue that Playwright starts up more slowly than Codex, so they’re not synchronized

how can I do that? If I just run the command in a terminal, codex doesn't pick it up. I'm using the vs code extension atm

edit: actually, same thing happens in the codex CLI upon startup:

🖐  MCP client for supabase failed to start: program not found

@sloikod

Before starting Codex, run:

npx playwright test

or, for a specific browser:

npx playwright open chromium https://example.com

npx playwright open firefox https://example.com

After that, when you launch Codex, Playwright should not throw an error and will start correctly.

I don't have access to a computer right now to confirm this precisely, but that's the general idea.

Playwright will simply start faster the second time, and there won’t be an error caused by incorrect synchronization.

@sloikod Before starting Codex, run: npx playwright test or, for a specific browser: npx playwright open chromium https://example.com npx playwright open firefox https://example.com After that, when you launch Codex, Playwright should not throw an error and will start correctly. I don't have access to a computer right now to confirm this precisely, but that's the general idea.

doesnt work for me, i can open example.com but codex wont find the mcp server.

>_ You are using OpenAI Codex in C:\Users\...

 To get started, describe a task or try one of these commands:

 /init - create an AGENTS.md file with instructions for Codex
 /status - show current session configuration and token usage
 /approvals - choose what Codex can do without approval
 /model - choose what model and reasoning effort to use

🖐  MCP client for `playwright` failed to start: program not found       

@SphaeroX
First, install Playwright locally. Then check on GPT how to do it

If you still can’t get it working, I’ll upload it in a few hours once I’m back at my PC."

I have Playwright installed, and I also installed Microsoft’s MCP server. I can run Playwright normally without any issues, but whenever I start Codex, I keep getting this error message. I don’t think this is just happening to me, other people using Windows are running into the same problem as well.

For example, with Kilo Code I don’t have any issues. Even Microsoft confirmed on GitHub that this isn’t a problem with the MCP server itself, but rather an issue with Codex.

Take a look at this https://github.com/openai/codex/issues/2905

Take a look at this #2905

Yes, thank you very much, but I would still prefer if the bug could be fixed directly in Codex before I start trying to work around the problem. It would make more sense to address the root cause first.

@SphaeroX Yeah, I hope they can fix it directly in the update. I'm not a developer either, so I can't submit code.

I have the same Problem under win 11 - no MCP Server would like to start :-( ! I have installed codex-cli 0.27.0 !!

<img width="726" height="455" alt="Image" src="https://github.com/user-attachments/assets/cc7906cc-46a6-4be0-b8ab-c6aab0eeca3f" />

@SphaeroX @jtuttas I wrote the reason why.
Each of the above will work if you run it before Codex.

<img width="630" height="270" alt="Image" src="https://github.com/user-attachments/assets/241cf4f4-499a-4f71-bbf8-5789c80be190" />

npx @playwright/mcp@latest --browser=chromium


<img width="1143" height="270" alt="Image" src="https://github.com/user-attachments/assets/1a81c09d-d430-4bc7-a26d-5d89fece14e2" />

@AdsQnn

are you on windows?

@AdsQnn are you on windows?

I’m using Ubuntu, but it doesn’t matter.
It’s just that Codex starts before MCP is loaded. You see that after the initial call, it loads fine.

Yes I'am on Win 11

~~It looks like only Windows is affected. As mentioned, it works with other tools, but not with Codex. I also started the server before~~

It looks like only Windows is affected. As mentioned, it works with other tools, but not with Codex. I also started the server before.

I clearly wrote that I’m on Ubuntu and even shared screenshots showing that the problem also occurs and how to solve it :) Please read carefully.

@AdsQnn I'm sorry if I'm mistaken, but are you not facing a different error? Yours says request timed out whilst ours is saying program not found. We are also using Windows, which this issue appears to affect specifically, though I'm not 100% sure. I think @SphaeroX was right

@sloikod
Does running npx @playwright/mcp@latest --browser=chromium directly in the terminal work correctly? After running this and then starting Codex, do you still experience the same problem?

Either way, it’s probably not a coincidence that the MCP loading error appears on both systems. Try adding the Codex MCP as shown in the screenshot and see if it starts — on my side, that’s the only one that runs without an error. So it’s clearly something related to the runtime.

In the screenshot @jtuttas you can see that the GitHub MCP also has a request timeout.

No need to apologize — everyone can make mistakes :) The point is just to get the problem solved

Besides, it also seems that on Windows the terminal may open in an environment where not all PATH variables propagate correctly. A similar issue occurred earlier in the context of env.

@sloikod

Sure, no problem. I just want to get it running here and have the issue fixed, not only for me but for others as well. I ran the command and nothing happened. There was no error message. I am not sure if that is expected, but when I try to launch Codex in a new tab afterward I get the same error again.

https://github.com/user-attachments/assets/05e98753-7653-4165-bfce-c38ebd35b99a

@AdsQnn I'm sorry if I'm mistaken, but are you not facing a different error? Yours says request timed out whilst ours is saying program not found. We are also using Windows, which this issue appears to affect specifically, though I'm not 100% sure. I think @SphaeroX was right

Same on VSCode, i started the Server and its running.

https://github.com/user-attachments/assets/5910f659-ae18-4929-ac5f-9a432f260f4b

Problem:

On Windows, Rust’s std::process::Command directly calls the system API (CreateProcessW).

This API does not append extensions (.exe, .bat, .cmd) and does not interpret batch scripts.

CMD/PowerShell perform that “magic” (PATHEXT resolution, batch handling), so something that works in the console fails in Rust with program not found.

MCP (client/server) tries to launch a process but cannot find the file in the required form → error.

Solution:

1. Provide the full filename with .exe or implement PATHEXT logic yourself (try .exe, .cmd, .bat).

1. For batch files (.bat, .cmd), launch them via cmd.exe /C ....

1. Use absolute paths and pass arguments with Command::arg(), not as a single string.

1. Set STDIO to piped, since MCP communicates over JSON-RPC through STDIN/STDOUT.

1. Build a native Windows binary (.exe) and place it in the directory the VS Code extension expects (…\bin\windows-x86_64\).

In short: Rust is not a shell, so you must explicitly specify the correct executable file and how to launch it. This ensures MCP starts correctly.

@SphaeroX
So you were right — the issue is indeed about program not found on Windows :)
However, fixing it requires a PR to Codex.

@AdsQnn
Yes, I can now confirm that the MCP server is running. I tested it with MCP Inspector and everything works. I hope this gets fixed soon for Windows Users...

https://github.com/user-attachments/assets/870accbc-afb0-420c-84a5-762eef76150e

I've looked at another issue and actually got it to now display the request timed out error:

MCP client for supabase failed to start: request timed out

Here's how I did it:

I found the node.exe executable, for me it's this: "C:\\Users\\username\\AppData\\Local\\nvm\\v22.14.0\\node.exe"

And I also ran this command in a separate terminal: npx -y @supabase/mcp-server-supabase@latest --access-token my_access_token

Then I manually searched through the C:\Users\username\AppData\Local\npm-cache\_npx directory to find the process that I just launched in the terminal and found the \dist\index.js file location, for me it was C:\\Users\\username\\AppData\\Local\\npm-cache\\_npx\\53c4795544aaa350\\node_modules\\@supabase\\mcp-server-supabase\\dist\\index.js

Then I changed config.toml to this:

[mcp_servers.supabase]
command = "C:\\Users\\username\\AppData\\Local\\nvm\\v22.14.0\\node.exe"
args = ["C:\\Users\\username\\AppData\\Local\\npm-cache\\_npx\\53c4795544aaa350\\node_modules\\@supabase\\mcp-server-supabase\\dist\\index.js"]

And now I'm getting the request timed out that @AdsQnn was talking about earlier. Not sure what to do now

Try:

npx -y @supabase/mcp-server-supabase@latest --access-token my_access_token

Alternatively, you can run it this way:

C:\Users\your_username\AppData\Local\nvm\v22.14.0\node.exe C:\Users\d\AppData\Local\npm-cache\_npx\53c4795544aaa350\node_modules\@supabase\mcp-server-supabase\dist\index.js

This will start the MCP server in a separate terminal. Keep that terminal open and running, and then in a different window, launch Codex.

I’m not entirely sure how this will behave in the Windows context, so please try it out and let me know what happens. The important thing is: keep the MCP server running in one terminal, while running Codex in another.

It’s probably caused by the fact that Codex starts before the MCP server is up — in other words, a synchronization issue. On my Ubuntu setup, it works.
PS.
@sloikod "I've looked at another issue and actually got it to now display the request timed out error:"
That’s good — I think it’s worth mentioning that I already wrote this above, since it outlines what the problem actually stems from.

Well so what I tried is: 'npx -y @supabase/mcp-server-supabase@latest --access-token my_access_token' in a separate terminal, then launched codex with this in the config.toml:

[mcp_servers.supabase]
command = "C:\\Users\\username\\AppData\\Local\\nvm\\v22.14.0\\node.exe"
args = ["C:\\Users\\username\\AppData\\Local\\npm-cache\\_npx\\53c4795544aaa350\\node_modules\\@supabase\\mcp-server-supabase\\dist\\index.js"]

Maybe one of these steps is unnecessary? Stopping the terminal process still results in the 'MCP client for supabase failed to start: request timed out' error, and, well, removing from the config.toml would just make the error not even show up so no

Ok, so the problem must be with the communication between the Windows shell and the Rust command, since on Ubuntu both the PATH and the initial startup work fine.

well I might just setup WSL, since that's what worked for #2555

cóż, może po prostu skonfiguruję WSL, ponieważ to zadziałało w przypadku #2555

Good idea — however, that removes the symptom rather than the root cause. It would be better to fix it properly sooner or later, without relying on such workarounds.

#2968
I’ve submitted a PR with the fix — let’s see how it goes.
I haven’t tested it on Windows directly, but it should work as intended.
On Ubuntu, the request timeout issue has been resolved.

Got the same exact issue. Tried debugging the whole day yesterday. No luck

I use the PlayWright MCP in Windows too,this is my mcp settings in config.toml
[mcp_servers.playwright]
command = "npx"
args = ["@playwright/mcp@latest"]
env = {}

I have the same problem that the codex reports MCP client for playwright failed to start: program not found

Did you solve the problem？

I use the PlayWright MCP in Windows too,this is my mcp settings in config.toml [mcp_servers.playwright] command = "npx" args = ["@playwright/mcp@latest"] env = {} I have the same problem that the codex reports MCP client for playwright failed to start: program not found Did you solve the problem？

It's not just for playwright, it's for every MCP.

ok after debugging and trying a ton of stuff - I fixed it. so I re-installed playwright and playwright-mcp globally for the quadruple time then just set my config.toml as follows

• [mcp_servers.playwright]
• command = "C:\\Users\\Your_User\\AppData\\Roaming\\npm\\playwright-mcp.cmd"
• args = []
• env = { MCP_DEBUG_PORT = "7110" }

calling the cmd directly seems to be just as fast as running codex on windows which does not provoke the race in between the two.

FYI -- Got Codex / Playwright MCP working on Windows11 w/o installing either globally.
Verify 'npx' path and replace 'xxxxxx' with your User name.

-- command fixes the 'Program Not Found'
-- env config resolved timeout.

[mcp_servers.playwright]
command = "C:\\\Users\\\xxxxxx\\\AppData\\\Roaming\\\npm\\\npx.cmd"
args = ["@playwright/mcp@latest"]

[mcp_servers.playwright.env]
SYSTEMROOT = 'C:\Windows'

This seems to also be an issue for WSL. 🥲

Yes, I sent a PR that uses which, which means you don’t have to manually replace it with the full path — it happens under the hood, making it Windows-friendly. This is required instead of running the Rust binary directly through a raw command.

I also added reactive loading instead of using a timeout for MCP.
Unfortunately, the PR hasn’t been reviewed at all, and now it already has conflicts.

---

Why which is needed in Rust:
The which crate in Rust locates the full path to an executable by searching through the system’s PATH environment variable. This ensures your program works reliably across platforms, especially on Windows, where direct execution often fails without resolving the full path first.

So, what @mikekidder does, which does automatically.

Why Windows has a problem with this:
On Windows, executables often can’t be run directly by name because the system doesn’t automatically resolve them through the PATH like Unix-based systems do. You need the full path to the binary, and which takes care of finding it for you.

Another way to fix for anyone still struggling with this is to install the MCP node package globally, then configure it like so:

[mcp_servers.github]
command = "C:\\Program Files\\nodejs\\node.exe"
args = ["C:\\Users\\[user_name]\\AppData\\Roaming\\npm\\node_modules\\@modelcontextprotocol\\server-github\\dist\\index.js", "--stdio"]
env = { "GITHUB_TOKEN" = "$env:GITHUB_TOKEN" }

This is obviously for github, so for Playwright you can remove the env section entirely, but swap out @modelcontextprotocol\\server-github for whatever package you've been trying to run via npx.

Initially I tried using the absolute path to the relevant .cmd file in the AppData\\Roaming\\npm directory, both with an empty args array and with args = ["--stdio"] but this always resulted in a request timed out error rather than the original program not found error. The key was using node.exe as the command and passing the arguments as above.

I have tested the following, but unfortunately it doesn’t work. And I can’t seem to get it running with any of the other setups either. Especially context7 would be another MCP that could be really useful.

[mcp_servers.playwright]
command = "C:\\Program Files\\nodejs\\npx.cmd"
args = ["@playwright/mcp@latest"]

@SphaeroX you should really just try WSL at this point. Switch took me like half an hour and I don't regret it. Though the issue on windows has to be fixed, I don't disagree

I tagged a thread somewhere earlier that had a guy explain the full setup in a single message, if you need it

You can try next toml config, it works for me under windows 11:

[mcp_servers.playwright-mcp]
command = "C:\\Program Files\\nodejs\\npx.cmd"
args = ["-y", "@playwright/mcp@latest", "--extension", "--browser", "chrome", "--executable-path", "C:\\Program Files\\Google\\Chrome\\Application\\chrome.exe"]
[mcp_servers.playwright-mcp.env]
SYSTEMROOT = 'C:\Windows'

to install extension, use https://github.com/microsoft/playwright-mcp/blob/main/extension/README.md

Inspired by this discussion, I've opened PR #3828 to resolve this issue.
@bolinfest - would appreciate a review when you have a chance. Thanks!

Just tested right now. The issue persists on Windows 11.

Having the same problem, I tried everything listed here, under Windows 10, nothing works, full path to npx.cmd does not work either. Sometimes it is hard to believe that a gazilion $ companies can not fix a problem that simple. There is an actual PR with a fix open, which is not moving... I will stick to Claude Code until this is fixed.

@n-kostadinov

Try this workaround:

1. Find the correct npx path: Run cmd /C where npx to list all npx locations. Test each path ending in .cmd as your command until one works.

1. Set environment variables: Add PATH, APPDATA, and SYSTEMROOT to your MCP server's env configuration. Get these values by running:

• cmd /C echo %PATH%
• cmd /C echo %APPDATA%
• cmd /C echo %SYSTEMROOT%

Important: When adding paths to the config, replace all single backslashes (\) with double backslashes (\\).

Example configuration:

[mcp_servers.playwright]
command = "C:\\Users\\YourUsername\\AppData\\Roaming\\npm\\npx.cmd"
args = ["-y", "@playwright/mcp@latest"]

[mcp_servers.playwright.env]
PATH = "C:\\Program Files\\nodejs\\;C:\\Users\\YourUsername\\AppData\\Roaming\\npm;C:\\Windows\\system32;C:\\Windows"
APPDATA = "C:\\Users\\YourUsername\\AppData\\Roaming"
SYSTEMROOT = "C:\\Windows"

Note: The PATH can be long. You don't need to include the entire PATH - just ensure it contains:

• The Node.js installation directory (e.g., C:\\Program Files\\nodejs\\)
• The npm directory (e.g., C:\\Users\\YourUsername\\AppData\\Roaming\\npm)
• Basic system paths (C:\\Windows\\system32;C:\\Windows)

Very simple alternative to run playwright mcp in windows without WSL and figuring out paths...

Start and keep running MCP Server:

npx @playwright/mcp@latest --port 8931

Then in your codex configuration file config.toml simply add:

[mcp_servers.playwright]
url = "http://localhost:8931/mcp"

Then execute your codex command.
This assumes both MCP Server and MCP Client (codex) run on the same host, if not then replace localhost with actual host where MCP Server is running.