PkgxSAFE

A [Model Context Protocol] server that can run any Open Source tool via pkgx.

Usage

1. Install [pkgx] (or not, we download it for you if you don’t)
2. git clone https://github.com/pkgxdev/pkgx-mcp
3. Consume in an MCP client (see below for configuration examples)

Cursor/Claude Configuation

• ~/.cursor/mcp.json:
• ~/Library/Application Support/Claude/claude_desktop_config.json

{
"mcpServers": {
"pkgx": {
"command": "/path/to/pkgx-mcp/main.ts"
}
}
}

Or without pkgx, but with npx:

{
"mcpServers": {
"pkgx": {
"command": "npx",
"args": ["ts-node", "/path/to/pkgx-mcp/main.ts"]
}
}
}

MCP Inspector

npx @modelcontextprotocol/inspector ts-node ./main.ts

Suggested Prompts

Use pkgx to run a series of git commands and generate statistics about this
repository.

Security

Security? Hah.

But seriously, this is a dangerous tool. You are letting AI read any file on your computer†. We have the decency to run pkgx in a sandbox so at least a rogue AI can’t run rm -rf /. But we have not restricted reads because that seemed too limiting.

We need the MCP protocol to advance to allow the user to be prompted with UI to approve certain things that the AI cannot by itself do or something like that.

In the meantime USE WITH ABSURD AMOUNTS OF CAUTION.

† well, on macOS at least we prevent reads to ~/.ssh and ~/.aws, PRs welcome to exclude more and add Linux sandboxing.

[!IMPORTANT] The AI cannot write to your file system. On macOS anyway. On Linux we need you to PR that for us.

[Model Context Protocol]: https://github.com/modelcontextprotocol [pkgx]: https://pkgx.sh