MCP Servers Don't Work with NVM

[combdn]

Problem

When using NVM (Node Version Manager), the standard installation and usage instructions for MCP servers don't work. The app tries to use an incorrect Node and fails.

Workaround

1. Avoid npx, install packages globally.
2. Use absolute paths to both the Node executable and server script.

Example:

{  
  "mcpServers": {
    "puppeteer": {
      "command": "/Users/username/.nvm/versions/node/v22.11.0/bin/node",
      "args": [
        "/Users/username/.nvm/versions/node/v22.11.0/lib/node_modules/@modelcontextprotocol/server-puppeteer/dist/index.js"
      ]
    }
  }
}

OS: macOS 15.1.1 (24B91)

[mckaywrigley]

I also ran into issues with the recommended npx setup, and I was able to successfully reproduce this workaround - fixes it completely.

MCPs now install and attach correctly within Claude Desktop.

[lborgav]

This workaround works fine

[adam91holt]

I can confirm this is working for me too.

You may have a different node.js version installed so just do the following for puppeteer as an example

npm i -g @modelcontextprotocol/server-puppeteer
cd /Users/YOUR_USERNAME/.nvm/versions/node
ls
>>> v14.17.6        v16.19.0        v18.14.0        v18.20.4

Then I ended up with this and it works sweet!

{
  "mcpServers": {
    "puppeteer": {
      "command": "/Users/YOUR_USERNAME/.nvm/versions/node/v18.20.4/bin/node",
      "args": [
        "/Users/YOUR_USERNAME/.nvm/versions/node/v18.20.4/lib/node_modules/@modelcontextprotocol/server-puppeteer/dist/index.js"
      ]
    }
  }
}

[jspahrsummers]

The challenge with NVM is that it seems to mostly work by installing a very complicated shell function into your shell profile. It'd be great if we could figure out your preferred Node version via NVM, but I haven't found a good way to do that—open to ideas!

[jspahrsummers]

Cross-linking: see https://github.com/modelcontextprotocol/servers/issues/40 for Windows workaround steps as well

[lolochka]

Unfortunately, the issue is still reproducible on Mac with nvm installed.

OS: macOS 15.1.1 Claude: 0.7.5 Locally I have checked that all packages are available with inputted paths.

The SQlite server is working correctly, but tools for server-filesystem doesn't appear

My settings

{
  "mcpServers": {
    "sqlite": {
      "command": "uvx",
      "args": ["mcp-server-sqlite", "--db-path", "/Users/user/test.db"]
    },
    "filesystem": {
      "command": "/Users/user/.nvm/versions/node/v22.11.0/bin/node",
      "args": [
        "-y",
        "/Users/user/.nvm/versions/node/v22.11.0/lib/node_modules/@modelcontextprotocol/server-filesystem/dist/index.js",
        "/Users/user/claude"
      ]
    }
  }
}

[mlaugharn]

suggestion: the same way that there is the mcpServers field, can there also be an eg mcpShell field? Or maybe a shell or source key per individual server, eg "shell": "source /Users/user/mcp.sh" ? Might introduce 'too much' flexibility but it could make mcp more usable/transparent to tinker with.

[daaain]

I was struggling to get MCP working yesterday, the error messages in the logs were unhelpful, and the most confusing of all is that it just started working today. This might not be the repo for this request, but it would be really useful if it was explicily documented which shell Claude Desktop is using to start the MCP services.

To your question @jspahrsummers, wouldn't nvm alias default work?

[jspahrsummers]

Unfortunately no, the environment for GUI apps is completely different from what you see on the command line.

[daaain]

Unfortunately no, the environment for GUI apps is completely different from what you see on the command line.

Ah right, I had a look and to answer my own question, assuming that Claude Desktop's Electron implementation uses the MCP TS SDK, it's node:child_process spawn with shell: false.

[combdn]

In the meantime, I have migrated from NVM to mise, and now, if I start the Claude app from my shell (open -a Claude), it can correctly use node and npm (it didn’t work with NVM). That might be a workaround for some people with the same problem.

PS Apple seems to have removed an option to set up the environment for the GUI apps through the launch agents. But I'm not sure. I didn’t dive too deep into it.

[daaain]

I was going to say check if path.join(process.env.HOME, '.nvm/nvm-exec') exists and if so, use it, but then also realised that path won't be the same for Windows and the config JSON doesn't specify if the MCP server is Node or Python and using npx -y is just a convention, so can't assume anything.

Feels like the current implementation is in a bit of an uncanny valley between being very open for people to run whatever they want while also very limited by the GUI spawned non-shell process 🤔 I'm wondering if offering the option to either use bundled runtimes for simplicity or running in a full shell for flexibility would be useful? That's what Mac Git GUIs I've used have been doing to solve similar issues. Or WASM 😅

Or actually, maybe having a separate daemon process to make this easier, especially if there are plans to integrate Claude deeper for computer use?

Edit: actually, ~/.nvm/alias/default is a simple text file with the NVM default node version, so if it exists then it can be used to construct a platform dependent path to the Node binary.

BTW MCPs stopped working with my Claude Desktop again, the errors in logs suggest it might be a command argument concat issue?

mcp-server-webresearch.log:

ERROR: You must supply a command.

Execute binaries from npm packages.

  npx [options] <command>[@version] [command-arg]...
...

mcp-server-filesystem.log

command not found: /Users/dain/Desktop

With the config:

{
  "globalShortcut": "",
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/dain/Desktop"
      ]
    },
    "webresearch": {
      "command": "npx",
      "args": ["-y", "@mzxrai/mcp-webresearch"]
    }
  }
}

[lynnbright]

This workaround works for me. Thanks a lot! 🙏 Here's how I successfully installed the MCP servers.

Environment

• macOS 14.3
• Claude 0.7.5

Installation Steps

1. Check your Node.js version:

node -v
# Example output: v18.17.0

2. Clone and set up the repository:

git clone <repository-url>
cd <repository-name>
npm install
npm run build
npm link

Configure claude_desktop_config.json

Locate Required Paths

1. Find your Node.js executable path:

cd ~/.nvm/versions/node
ls  # List available Node versions
cd v18.17.0/bin  # Replace with your version
pwd  # Copy this path + /node for the command field
# Example output: /Users/lynnbright/.nvm/versions/node/v18.17.0/bin/node

2. Find the built file path:

cd ~/.nvm/versions/node/v18.17.0/lib/node_modules/linear-mcp-server/build
pwd  # Copy this path + /index.js for the args field
# Example output: /Users/lynnbright/.nvm/versions/node/v18.17.0/lib/node_modules/linear-mcp-server/build/index.js

Example Configuration

{
  "mcpServers": {
    "sqlite": {
      "command": "uvx",
      "args": [
        "mcp-server-sqlite",
        "--db-path",
        "/Users/lynnbright/test.db"
      ]
    },
    "linear": {
      "command": "/Users/lynnbright/.nvm/versions/node/v18.17.0/bin/node",
      "args": [
        "/Users/lynnbright/.nvm/versions/node/v18.17.0/lib/node_modules/linear-mcp-server/build/index.js"
      ],
      "env": {
        "LINEAR_API_KEY": "lin_api_xxxx"
      }
    },
    "notion": {
      "command": "/Users/lynnbright/.nvm/versions/node/v18.17.0/bin/node",
      "args": [
        "/Users/lynnbright/.nvm/versions/node/v18.17.0/lib/node_modules/notion/build/index.js"
      ],
      "env": {
        "NOTION_API_TOKEN": "ntn_xxxxx"
      }
    }
  }
}

Final Step

Restart Claude Desktop.

[laulauland]

FNM (https://github.com/Schniz/fnm) doesn't work either but the workaround worked!

[mearleycf]

In the meantime, I have migrated from NVM to mise, and now, if I start the Claude app from my shell (open -a Claude), it can correctly use node and npm (it didn’t work with NVM). That might be a workaround for some people with the same problem.

PS Apple seems to have removed an option to set up the environment for the GUI apps through the launch agents. But I'm not sure. I didn’t dive too deep into it.

This is the way. Mise is an infinitely better tool than NVM.

[gannonh]

Interesting. Btw, the NVM issue goes away if you're only managing 1 version of node. I discovered this inadvertently setting up a new machine. Will check out Mise. First I've heard of it.

[serverlesspolska]

This is the way. Mise is an infinitely better tool than NVM.

Even if it is. Nevertheless, nvm is very popular, and it would be nice if it were supported by Claude MCP.

[zaphodtx]

The challenge with NVM is that it seems to mostly work by installing a very complicated shell function into your shell profile. It'd be great if we could figure out your preferred Node version via NVM, but I haven't found a good way to do that—open to ideas!

Same goes for Volta (volta.sh).

[ycjcl868]

Same issue with fnm(https://github.com/Schniz/fnm) Node.js version manager, when using node or npx:

spawn node ENOENT
    at Process.ChildProcess._handle.onexit (node:internal/child_process:286:19)
    at onErrorNT (node:internal/child_process:484:16)
    at processTicksAndRejections (node:internal/process/task_queues:82:21) {
  errno: -2,
  code: 'ENOENT',
  syscall: 'spawn node',
  path: 'node',
  spawnargs: [Array]
}

[rgarcia]

Potentially easier workaround that I don't think has been mentioned yet--place a script somewhere like /usr/local/bin/npx-for-claude:

#!/bin/zsh
source ~/.zshrc
exec npx "$@"

Then chmod +x /usr/local/bin/npx-for-claude. Might need to riff on it if you use bash, but this works for me. Now I can do something like this in my config:

{
  "mcpServers": {
    "everything": {
      "command": "npx-for-claude",
      "args": ["-y", "@modelcontextprotocol/server-everything"]
    }
  }
}

Have only tested that this works with fnm. Not sure about nvm.

[Ranteck]

the issue still happening. I fix it doing this

      "command": "C:\\nvm4w\\nodejs\\node.exe",
      "args": [
        "C:/Users/USERNAME/Documents/MCPServer/node_modules/@modelcontextprotocol/server-memory/dist/index.js"
      ]
    },```

[T1T4N]

I followed @rgarcia's workaround and for me on macOS it was enough that the wrapper script just includes the following lines:

#!/usr/bin/env bash

export PATH="/opt/homebrew/bin:$PATH"
exec npx "$@"

Also tested the same wrapper script with pnpm from corepack, worked wonderfully.

[TomExMachina]

Try this one: https://github.com/ihor-sokoliuk/mcp-searxng

I can't for the life of me get it working with nvm. I tried the command proxying of node and npx, sourcing .zshrc, absolute paths, global install, manually building ahead of time.

[MarcusQuirino]

@rgarcia worked for me. macbook air M1 with FNM and zsh

[Ranteck]

@rgarcia worked for me. macbook air M1 with FNM and zsh

it works without problem in macbook but in windows is another thing

[rededge-ian]

For anyone who is using Volta for node version management, the workaround in the OP should work when paired with volta which:

1. First, change the command from npx and instead use the full path to node, which you can find by running volta which node.
2. Install the mcp package globally (e.g. npm install -g figma-mcp)
3. Run volta which [package] to get their real path to use as the first argument under args. This will result in something like the following:

{
  "mcpServers": {
    "figma-mcp": {
      "command": "/Users/username/.volta/tools/image/node/20.18.3/bin/node",
      "args": [
        "/Users/username/.volta/tools/image/packages/figma-mcp/bin/figma-mcp"
      ],
      "env": {
        "FIGMA_API_KEY": ""
      }
    }
  }
}

(Note that if you did not want to have to update this for each new version of node, for step 1 you could instead use a variant of @T1T4N's solution with a shell script which sets the VOLTA_HOME and PATH env variables to the correct paths.)

[intertwine]

For Volta on OS X (15.4) and Claude Desktop (v0.8.0), using the full path to Volta's npx resolved my MCP connection timeout errors

In ~/Library/Application\ Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "filesystem": {
      "command": "/Users/username/.volta/bin/npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/username/dirA/",
        "/Users/username/dirB/"
      ]
    },
    "brave-search": {
      "command": "/Users/username/.volta/bin/npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"],
      "env": {
        "BRAVE_API_KEY": "<API_KEY>"
      }
    }
  }
}

[renanwilliam]

There is no way to make the @rgarcia work using an Intel Mac. But using the workaround of installing packages globally works fine.

[drewjs]

#!/usr/bin/env bash

fnm exec --using=default npx "$@"

If you're using fnm the above script worked well for my setup. Just add to "/usr/local/bin" and make sure you chmod +x accordingly. You can replace "default" with a specific version (eg. --using=20) if you want to always run MCP servers with a specific version of node/npm

[riemannzeta]

Confirming that this issue and the reported workaround still exist as of today with Claude Desktop 0.8.0 running on macOS Sequoia 15.3.1.

Using the claude_desktop_config.json suggested in the walkthrough results in the following errors in mcp-server-filesystem.log:

npm
 ERR! Invalid dependency type requested: alias

npm ERR! A complete log of this run can be found in:
npm ERR!     /Users/mfmartin/.npm/_logs/2025-03-09T18_51_33_223Z-debug.log
Install for @modelcontextprotocol/server-filesystem@latest failed with code 1

and the content of /Users/mfmartin/.npm/_logs/2025-03-09T18_51_33_223Z-debug.log shows that Claude Desktop is defaulting to an old version of node when it attempts to install @modelcontextprotocol/server-filesystem:

1 verbose cli [ '/Users/mfmartin/.nvm/versions/node/v10.13.0/bin/node',
1 verbose cli   '/Users/mfmartin/.nvm/versions/node/v10.13.0/lib/node_modules/npm/bin/npm-cli.js',
1 verbose cli   'install',
1 verbose cli   '@modelcontextprotocol/server-filesystem@latest',

When run from command line, npx works fine:

mfmartin@m1a ~ % npx -y @modelcontextprotocol/server-filesystem /Users/mfmartin/Desktop /Users/mfmartin/Downloads
Secure MCP Filesystem Server running on stdio
Allowed directories: [ '/Users/mfmartin/Desktop', '/Users/mfmartin/Downloads' ]

Presumably working fine because it's running off the nvm default alias 23.9.0. I tried in vain to figure out how to override whatever Claude Desktop was calling with environment variables (like NODE_PATH) with weird intermittent success.

But the reported workaround of running the server using node and the server index.js from an absolute path resolved the error:

{
  "mcpServers": { 
	"filesystem": {
	  "command": "/Users/mfmartin/.nvm/versions/node/v23.9.0/bin/node",
	  "args": [
	    "/Users/mfmartin/.nvm/versions/node/v23.9.0/lib/node_modules/@modelcontextprotocol/server-filesystem/dist/index.js",
	    "/Users/mfmartin/Desktop",
	    "/Users/mfmartin/Downloads"
	  ]
	}
  }
}