RunOnMine
GitHub
MCP security
MCP makes it possible for AI applications to call external tools. RunOnMine focuses on the security boundary around those calls when the tools can read files, execute commands, drive a browser or control a desktop.
MCP security is the set of controls that determine which client may invoke which tool, against which resource, under which authority. RunOnMine evaluates those decisions on the target machine instead of treating connection to an MCP server as blanket permission.
Filesystem tools are scoped to directories the owner explicitly selects. Choosing a project root is intentionally narrower than granting an AI process the entire home directory.
Requests that need approval stop locally. Approval is bound to the connector, requester identity, tool and argument hash, so changing the concrete action can require a new decision.
Remote connectors cannot approve their own dangerous requests and cannot gain remote administrator execution. A broader local policy does not remove that remote ceiling.
The local owner can stop the agent and managed connectors, reject queued approvals, revoke OAuth sessions and invalidate temporary connector credentials with one explicit lock action.