How to Use AI Agents With Local Files via MCP on Mac
To let an AI agent work on local files through MCP, run a local MCP server that exposes specific tools, approve only the folders it may use, and have it write results as new files. Claude Desktop, Claude Code, Cursor and Codex can start such a server on your Mac. Check what the agent itself reads: that content can reach its model provider.
What MCP is, in one paragraph
The Model Context Protocol is a standard that lets an AI assistant call tools you provide. A local server speaks it over standard input and output, so the assistant starts a command on your machine and sends it requests.
Safety rules worth insisting on
Approve folders explicitly, never your whole home folder. Refuse access to credential folders such as .ssh. Write outputs as new files and never overwrite. Offer no delete and no network tools. Those limits make a mistake recoverable.
How Paperlane for Mac does it
Paperlane for Mac is a free native macOS app that lets approved AI clients run file tools on your Mac. You switch on agent access in Settings, approve folders, and pick a client. The helper forwards calls to the running app through a private socket. Agents can list tools, run them, check progress and cancel. Agent access is free, like the rest of the app.
Limits
The model provider can see anything the agent reads, so approve only folders you are happy to share with that agent. Clients that only support remote servers cannot use a local helper.
Frequently asked questions
Does MCP upload my files?
Paperlane for Mac’s tools do not send file bytes to the agent. Content the agent itself opens and reads can go to its model provider, so approve folders carefully.
Can a remote-only MCP client use Paperlane for Mac?
No. The client must be able to launch a local stdio command; the helper connects to the app through a private local socket.