MCP Configuration is not working

[ashfaq92]

Thanks for the package. I am having issues with MCP. It still says that the No servers added even though my /home/user_name/.config/Claude/claude_desktop_config.json file has:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/home/yourusername/Desktop",
        "/home/yourusername/Downloads"
      ]
    },
     "wot": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "http://localhost:3000/mcp",
        "--allow-http"
      ]
    }
  }
}

Please see the image below:

Specs:
Description: Ubuntu 24.04.3 LTS
Release: 24.04
Codename: noble

[eon01]

Logs? Typically:

• ~/.config/Claude/logs/mcp-server-filesystem.log
• ~/.config/Claude/logs/mcp-server-wot.log
• ~/.config/Claude/logs/mcp.log

or mcp-server-<name>.log in general.

[aaddrick]

Check ~/.cache/claude-desktop-debian too

[dr-msr]

Had the same issue, apparently the MCP configurations in the claude_desktop_config.json gets overwritten/reset easily. I had to maintain a copy of backup config so that whenever it got overwritten i can restore it easily.

[aaddrick]

Flagging #238 as a duplicate of this — same symptom on Fedora/DNF, closing it and tracking here.

I dug into the app code. Our packaging doesn't touch any MCP code paths, so this isn't something we introduced. The behavior is upstream.

Here's what I think is happening — two plausible explanations:

Most likely: npx isn't in PATH when Electron spawns child processes. The app loads your mcpServers config fine, tries to start the servers, they all fail immediately because npx resolves from your shell but not from Electron's environment. The UI then shows the MCP panel empty because all servers are in a failed state. The config file itself is never written.

Second possibility: Zod schema validation silently dropping the block. The app validates MCP entries against a schema that requires command to be a string. If any entry fails validation, the whole mcpServers block may be silently dropped on read. This would look identical from the outside.

The read-only file test in #238 actually supports the first explanation — if the file can't be written but behavior is the same, the file probably isn't being rewritten at all. The servers are just failing to start.

To narrow it down, a few things would help:

1. Full content of ~/.config/Claude/claude_desktop_config.json — paste the whole thing (redact paths if needed). I want to confirm the exact structure, especially that command is present on every entry.

2. Does the file actually change? Run cp ~/.config/Claude/claude_desktop_config.json /tmp/config-before.json, launch the app, then diff /tmp/config-before.json ~/.config/Claude/claude_desktop_config.json. Paste the output.

3. App logs — ~/.config/Claude/logs/main.log and any mcp-server-*.log files in that directory. Even if they're empty that's useful.

4. npx check — run this from a terminal and paste the output:

   which npx && npx --version
   

The npx PATH issue is the most common cause of this exact symptom, and it's something we could address on our end if that's the culprit.

---

Written by Claude Sonnet 4.6 via Claude Code

[joeyzapp]

Confirmed: npx PATH issue on Fedora 43 Workstation

Your first theory was correct — Electron wasn't finding npx even though it's available in the shell.

Environment

• OS: Fedora 43 Workstation (Wayland, running via XWayland)
• Install method: DNF repo
• Claude Desktop: v1.1.3189
• Node: v24.13.0
• npx: v10.9.4 at /usr/bin/npx

What didn't work

Config with bare npx:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/home/user/Documents/vault"
      ]
    }
  }
}

• No MCP-related entries appeared in main.log at all
• No mcp.log or mcp-server-*.log files were created
• The config file was not overwritten by the app (rules out schema validation dropping the block)
• The connectors menu showed nothing

What fixed it

Changing "command" to the full path:

{
  "mcpServers": {
    "filesystem": {
      "command": "/usr/bin/npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/home/user/Documents/vault"
      ]
    }
  }
}

After this change, mcp.log showed the server initializing and connecting successfully, and mcp-server-filesystem.log confirmed tools were registered. The filesystem connector appeared in the UI under Plus → Connectors and worked as expected.

Additional notes

• killall claude-desktop doesn't kill the app — the process name is electron, not claude-desktop. Use pkill -9 -f "electron.*claude-desktop" instead.
• Stale SingletonLock/SingletonSocket/SingletonCookie files in ~/.config/Claude/ can cause the app to think it's a second instance ("Not main instance, returning early from app ready" in main.log). Deleting those and doing a clean restart fixes it.
• The "Starting app" line sometimes appears twice in main.log even on a clean single launch — not sure if that's intentional.

Suggestion

Could the launcher script or the app's MCP spawn logic be updated to resolve npx (and node) against common paths like /usr/bin/ on Linux? This would save Fedora/RHEL users from having to figure out the full path workaround.

[pm64]

I am seeing the same behavior, in which claude_desktop_config.json is "reset" each time the Chat/Cowork/Code tab is changed (including on launch), despite having the full path to the MCP server command specified.