spawn npx ENOENT filesystem MCP

Not specified by the original author.

Fix npm error could not determine executable to run for exa-mcp-server

TL;DR

Point your MCP config at the server file with node directly instead of npx, or switch to the hosted remote server at https://mcp.exa.ai/mcp. The npx wrapper cannot resolve the package's entry point because the published bin field points inside a hidden .smithery directory, which npm refuses to execute.

The exact error:

npm error could not determine executable to run

Steps

1. Confirm this is the error

Run the package by hand in a terminal:

npx -y exa-mcp-server

Success check: if you see npm error could not determine executable to run, this skill applies. If instead you see a missing API key error, that is a different problem.

2. Find where npm installed the package

npm root -g

Success check: it prints a directory. Append /exa-mcp-server/.smithery/stdio/index.cjs to it. Verify the file exists:

ls "$(npm root -g)/exa-mcp-server/.smithery/stdio/index.cjs"

3. Point the MCP config at node directly

Edit claude_desktop_config.json (or your client's MCP config) and replace the npx entry with a direct node call:

{
  "mcpServers": {
    "exa": {
      "command": "node",
      "args": ["[global npm root]/exa-mcp-server/.smithery/stdio/index.cjs"],
      "env": { "EXA_API_KEY": "[your Exa API key]" }
    }
  }
}

Success check: save the file, fully restart the client. The exa server should start and its tools (web search, code context) appear in the tool list.

4. Alternative: use the hosted remote server instead

If you would rather skip the local install entirely, use the remote endpoint. For Claude Code:

claude mcp add exa -e EXA_API_KEY [your value] Exa API key] -- npx -y mcp-remote https://mcp.exa.ai/mcp

Success check: claude mcp list shows exa connected, and no local package install is needed.

When this applies

• You followed the standard npx -y exa-mcp-server setup for Claude Desktop, Cursor, or another MCP client and the server never starts.
• The terminal reproduces npm error could not determine executable to run for the package.

When it does not apply

• You get EXA_API_KEY environment variable is not set instead. That is an auth problem: put your key in the env block.
• You are using the remote https://mcp.exa.ai/mcp URL already. Do not mix the remote URL and the npx package in one entry; they are separate setups.

Tool compatibility

• exa-mcp-server npm package, versions publishing the .smithery/stdio/index.cjs layout
• npx / npm (Node.js 18+)
• Claude Desktop, Claude Code, Cursor, KiloCode

Why it happens

The published package's bin field points at .smithery/stdio/index.cjs, a path inside a hidden dot-directory. npm/npx will not resolve an executable out of a hidden directory, so it exits with the "could not determine executable" error even though the file is present and valid.

Edge cases

• bunx exa-mcp-server hits a related startup crash (SyntaxError at line 2 of the same file) on some versions; the node-direct workaround avoids that path too.
• If npm installed the package in a user-local prefix (.npm-global), the args path must use that prefix, not the system one. npm root -g tells you which.
• Some users report success pinning an older package version (e.g. [recipient email]) as a stopgap; the remote server option is the cleaner long-term move.

A user following the standard local stdio setup for the Exa MCP server ran npx -y exa-mcp-server and hit npm error could not determine executable to run. The issue thread diagnosed that the package's bin field points into the hidden .smithery directory, which npx cannot resolve. The verified workarounds from the thread: invoke node directly on the installed .smithery/stdio/index.cjs path in the MCP config, or drop the local package and use the hosted remote server https://mcp.exa.ai/mcp (e.g. via mcp-remote with the API key passed as EXAAPIKEY). Source: https://github.com/exa-labs/exa-mcp-server/issues/107

Not specified by the original author.

TL;DR: uvx is not a system command, it ships with the uv package manager, and the SQLite MCP server needs it to launch. Install uv, confirm uvx --version works in a terminal, then restart your MCP client.

spawn uvx ENOENT

Variant: uvx: command not found.

Fix it

1. Install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

On Windows: powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex".

2. Verify it is on PATH in a fresh terminal:

uvx --version

Expected: a version number like uv 0.7.x.

3. If the terminal sees it but the MCP client does not, the client was launched before uv updated PATH, or GUI apps do not read your shell profile. Either restart the machine, or put the absolute path to uvx in the client config "command" field. Find it with which uvx.

4. Restart the MCP client.

   Expected: the sqlite server spawns and its tools appear.

When to use this

• The MCP client shows the sqlite server as failed/disconnected immediately at startup.
• Logs or errors mention uvx, ENOENT, or command not found.

When NOT to use this

• The server starts but queries fail. That is a database problem, not a launcher problem.
• You installed the server via pip or another method. Then uvx is not involved at all.

Compatibility

• @modelcontextprotocol/server-sqlite installed via uvx.
• Any MCP client that spawns stdio servers: Claude Desktop, Claude Code, Cursor.

Why it happens

uvx is uv's tool-runner, similar to npx for npm. The official install docs assume it exists, but fresh machines (and especially GUI-launched MCP clients with a minimal PATH) do not have it. The client then fails to spawn the server process at all, which surfaces as ENOENT rather than any database error.

Edge cases

• Homebrew installs of uv usually land on PATH for terminals but not for GUI apps. The absolute-path trick in step 3 is the reliable fix on macOS.
• Corporate machines sometimes block the install script. pip install uv or brew install uv are alternatives.
• After installing, the first uvx run downloads the server package. Give it a minute on slow networks.

The official @modelcontextprotocol/server-sqlite is installed and run via uvx (uv run --with mcp-server-sqlite), which requires the uv package manager from Astral. The spawn uvx ENOENT failure is the standard symptom when uv is not installed or not on PATH, and the documented fix is installing uv from https://docs.astral.sh/uv/ before configuring the server. This matches the widely reported setup failure for uvx-based MCP servers.

Not specified by the original author.

TL;DR

It looks like npx writes install logs to stdout, which breaks the JSON-RPC handshake. The MCP protocol needs stdout reserved for JSON-RPC, and npx sometimes prints things like Need to install the following packages to stdout, which kills the handshake with an EOF error.

Steps

1. The recommended npx -y @perplexity-ai/mcp-server config fails in strict MCP clients with an initialize: EOF error. It looks like npx writes install logs to stdout, which breaks the JSON-RPC handshake. This happens consistently, not intermittently. The same request works in the playground but fails through the API with my key.
2. You diagnosed it right. The MCP protocol needs stdout reserved for JSON-RPC, and npx sometimes prints things like Need to install the following packages to stdout, which kills the handshake with an EOF error. The reliable fix is to skip the npx wrapper: install the package locally with npm and point your client config straight at the node executable and the server script. This is also documented in the repo troubleshooting section with the -yq flag for clients that hit it. Source: https://github.com/perplexityai/modelcontextprotocol/issues/74

When to use

You are seeing this: The recommended npx -y @perplexity-ai/mcp-server config fails in strict MCP clients with an initialize: EOF error. Use this skill when you run into "Perplexity MCP npx wrapper causes EOF errors in strict MCP clients".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Why this happens

It looks like npx writes install logs to stdout, which breaks the JSON-RPC handshake.

Sources

• https://github.com/perplexityai/modelcontextprotocol/issues/74

Someone asked: The recommended npx -y @perplexity-ai/mcp-server config fails in strict MCP clients with an initialize: EOF error. It looks like npx writes install logs to stdout, which breaks the JSON-RPC handshake. This happens consistently, not intermittently. The same request works in the playground but fails through the API with my key. The verified resolution was: You diagnosed it right. The MCP protocol needs stdout reserved for JSON-RPC, and npx sometimes prints things like Need to install the following packages to stdout, which kills the handshake with an EOF error. The reliable fix is to skip the npx wrapper: install the package locally with npm and point your client config straight at the node executable and the server script. This is also documented in the repo troubleshooting section with the -yq flag for clients that hit it. Source: https://github.com/perplexityai/modelcontextprotocol/issues/74 Original thread: https://vectle.com/threads/thrc84JuHMD5PwfcG9EcJS8A