Dear Authors / Maintainers,
We are researchers from the University of St. Gallen studying how to make Model Context Protocol (MCP) servers safer to run via a sandboxed permission system. As part of our study, we auto generated a permission manifest for your MCP server and would love your feedback on whether it is correct and complete.
The MCP server in question is: Clerk
Please review the manifest below and let us know:
- Are the permissions and their scopes correct?
- Are any permissions missing?
- Do any permissions need to be runtime-scoped (e.g., a specific project directory) rather than global?
Proposed manifest (please review)
{
"description": "Clerk Agent Toolkit MCP server: exposes Clerk backend functionality (users, organizations, invitations) as MCP tools over stdio. Authenticates via environment variables and performs outgoing HTTPS requests to Clerk APIs.",
"permissions": [
"mcp.ac.system.env.read",
"mcp.ac.network.client"
]
}
Please let us know if you have any questions and/or remarks.
In case you want to see the (current) full permission system:
MCP Permission System
| Permission |
Description |
Notes |
mcp.ac.filesystem.read |
Read files/directories |
|
mcp.ac.filesystem.write |
Write/create files |
|
mcp.ac.filesystem.delete |
Delete files or directories |
|
mcp.ac.system.env.read |
Read environment variables |
e.g., API_KEY, PATH |
mcp.ac.system.env.write |
Set environment variables |
setting the env variables |
mcp.ac.system.exec |
Execute OS commands |
CLI runners, shells |
mcp.ac.system.process |
List or kill processes |
|
mcp.ac.network.client |
General Outgoing network access |
|
mcp.ac.network.server |
Accept incoming connections |
|
mcp.ac.network.bluetooth |
Use Bluetooth connections |
macOS TCC-protected |
mcp.ac.peripheral.camera |
Capture images/video |
macOS TCC-controlled |
mcp.ac.peripheral.microphone |
Record audio |
TCC-protected |
mcp.ac.peripheral.speaker |
Play audio |
|
mcp.ac.peripheral.screen.capture |
Screen capture |
Requires consent (macOS: Screen Recording) |
mcp.ac.location |
Access location data |
From Wi-Fi, IP, GNSS |
mcp.ac.notifications.post |
Show system notifications |
macOS/Windows |
mcp.ac.clipboard.read / .write |
Read/write clipboard |
Copy-paste support |
Thank you very much for your time and your efforts in making MCP more secure.
Dear Authors / Maintainers,
We are researchers from the University of St. Gallen studying how to make Model Context Protocol (MCP) servers safer to run via a sandboxed permission system. As part of our study, we auto generated a permission manifest for your MCP server and would love your feedback on whether it is correct and complete.
The MCP server in question is: Clerk
Please review the manifest below and let us know:
Proposed manifest (please review)
{ "description": "Clerk Agent Toolkit MCP server: exposes Clerk backend functionality (users, organizations, invitations) as MCP tools over stdio. Authenticates via environment variables and performs outgoing HTTPS requests to Clerk APIs.", "permissions": [ "mcp.ac.system.env.read", "mcp.ac.network.client" ] }Please let us know if you have any questions and/or remarks.
In case you want to see the (current) full permission system:
MCP Permission System
mcp.ac.filesystem.readmcp.ac.filesystem.writemcp.ac.filesystem.deletemcp.ac.system.env.readAPI_KEY,PATHmcp.ac.system.env.writemcp.ac.system.execmcp.ac.system.processmcp.ac.network.clientmcp.ac.network.servermcp.ac.network.bluetoothmcp.ac.peripheral.cameramcp.ac.peripheral.microphonemcp.ac.peripheral.speakermcp.ac.peripheral.screen.capturemcp.ac.locationmcp.ac.notifications.postmcp.ac.clipboard.read/.writeThank you very much for your time and your efforts in making MCP more secure.