Codex Desktop launches Playwright MCP with cwd /, causing output path resolution to /.playwright-mcp

Same error here

I can reproduce the same issue on Desktop.

In my case there were two separate failures in sequence:

• first, a temporary Browser is already in use conflict from another Playwright session
• after closing that other session, the remaining failure was the same ENOENT: no such file or directory, mkdir '/.playwright-mcp'

The root cause looks like a generic stdio MCP cwd defaulting problem rather than something Playwright-specific:

• Playwright MCP falls back to process.cwd() when it does not get client roots / workspace context
• in the failing Desktop path, the stdio MCP child appears to start with cwd=/
• Playwright then resolves its output dir under /.playwright-mcp
• browser_navigate includes a snapshot in the response path, so it eventually tries to create that directory and fails there

The smallest fix seems to be on the Codex side when normalizing / preparing MCP server configs, not in Playwright itself:

• for McpServerTransportConfig::Stdio, if cwd is unset, default it to the active session/workspace cwd
• preserve explicitly configured cwd values as-is
• apply the same normalization on refresh/reconnect paths too, so the injected cwd is stable across session updates

Concretely, the relevant spot looks like the MCP server aggregation / normalization path in codex-rs/core/src/mcp/mod.rs (for example around effective_mcp_servers()), plus the session refresh path that rebuilds MCP state.

That should fix this for Playwright MCP and any other stdio MCP server that implicitly depends on the process working directory.

Worked around this issue. You can specify the --output-dir manually and override it trying to create a folder in the root level.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": [
        "@playwright/mcp@latest",
        "--output-dir=/Users/<your username>/.playwright-mcp"
      ]
    }
  }
}

Of course, the dir can be anywhere that is writable.

I've faced with the same issue.

Why does VSCode launch Codex with the current working directory set to /? What is the sense?

Confirmed this workaround fixes it for me in Codex Desktop as well.

I added --output-dir=/Users/hyejun/.playwright-mcp to the Playwright MCP args, restarted Codex, and the MCP was then able to launch and navigate successfully without trying to write to /.playwright-mcp.

So at least as a workaround, overriding the output dir to a writable location works.

same issue here
mkdir '/.playwright-mcp‘

very annoying bug (and I'm building my own MCP), specific to codex app (not to codex cli)

so the only work-around I discovered so far is to create .codex/config.toml in your project's directory and set cwd to the directory that MCP should use as work directory:

[mcp_servers.drymint]
# set to your project's directory
cwd = "/Users/me/TempProjects/project"
# other params
args = ["mcp", "run"]
startup_timeout_sec = 20.0

then at least you can set it up for your different projects (but obviously don't commit this)

+1, I am seeing the same Codex Desktop Playwright MCP behavior on macOS: screenshots try to write under /.playwright-mcp even though the workspace has a writable .playwright-mcp directory. The --output-dir workaround is needed here too.