Disclosure: I make Bridgerator, the tool this post is about. I'll stick to what it actually does and say where its limits are.
If you let an AI app work with files on your own Windows PC, the useful question is not "is it safe?" but "what happens when the AI makes a mistake?". Models misread instructions, rename the wrong folder or run a command you did not expect. Here is a short checklist of the settings that matter, using Bridgerator (an MCP bridge for ChatGPT, Claude, Codex, Cursor and other MCP apps) as the example. Most of it applies to any local MCP server.
1. Know the default
Bridgerator's defaults restrict nothing, like similar tools: with no folder list, the AI may work in every folder, except Bridgerator's own settings folders, which are always closed. If that is more access than you want, change it before the first real task, not after.
2. Limit the folders
Set an allowed-folders list, so the AI can only touch, say, D:\Projects and Documents\Invoices .…