Learn

How AI agents connect to a shared drive: MCP, CLI, API or a mounted drive



AI agents connect to a shared drive in one of four ways: through a mounted drive that looks like a normal folder on the computer, through an MCP server that the agent's app talks to, through a command-line tool, or through an API.

Which one to use depends on where the agent runs. An agent with access to your computer's files, such as Claude Code or Codex working in a terminal, can use the mounted drive. An assistant in a browser or phone app, such as Claude or ChatGPT, has no local folder to open, so it connects through MCP. Scripts and scheduled jobs usually call the CLI or the API.

Most teams use more than one, and all of them reach the same files, which is the point of a shared drive for AI agents.


The four routes compared

Route

Best for

What the agent sees

How it signs in

Mounted drive

Agents running on your computer (Claude Code, Codex, Cursor, local scripts)

A normal folder it can list, open and save to

Your desktop app session

MCP server

Chat assistants and AI tools without a filesystem (Claude, ChatGPT, Cursor, VS Code)

Tools for searching, reading files and creating notes, tasks, folders, tags and memories

OAuth, no API keys

CLI

Terminal work, shell scripts, coding agents in sandboxes

Commands it runs from a shell

See the CLI download page

API

Your own applications, automations and backends

REST endpoints

API key from workspace settings


A mounted drive for agents on your computer

The simplest route for local agents is to make the shared drive a folder. The Fabric desktop app mounts Fabric as a drive on Mac, Windows and Linux. Files take no space on your disk and stream when they're opened, so a folder holding years of client work or a large footage library doesn't fill up a laptop.

To an agent, the mounted drive is just another directory. Claude Code can list it, open brief.md, write plan.md and save results next to the source files, using the file tools it already has. Nothing needs configuring inside the agent, and the same goes for Cursor, a local Codex session or a Python script you run yourself.

There are two things worth knowing. The mounted drive only helps agents that run on that computer, so it does nothing for Claude on your phone or a colleague's ChatGPT. And opening files one by one through a folder is the slow way to search a very large drive. For broad searches, the agent is better off using Fabric's server-side search through MCP or the CLI, which runs grep, regular expressions and full-text search across every file without downloading them. Why AI agents give up searching your files explains why that matters.


The MCP server for Claude, ChatGPT, Cursor and coding agents

The Model Context Protocol is an open standard that lets an AI tool call outside services as tools. Fabric's MCP server lives at https://mcp.api.fabric.so/mcp. Sign-in is OAuth, so there are no API keys to copy or store, and setup takes under a minute. Once connected, the AI tool can search Fabric, read file contents (including files from apps you've connected to Fabric, such as Google Drive), and create or edit notes, bookmarks, folders, tasks, tags and memories.

Setup differs slightly in each tool.

Claude. Fabric is listed in Claude's official connector directory, so you can add it from there. To add it by hand in Claude Desktop, open Settings and go to Connectors, then choose to add a custom connector. Enter a name and the server URL, then follow the sign-in prompts.

ChatGPT. Open Settings and go to Connectors, then choose Add Connector. Enter the server URL, give it a name, and save. Then start a new conversation, click More, and choose Add Connectors. Select Fabric and sign in when prompted.

Cursor. Open Cursor Settings, go to MCP, and choose Add new MCP Server. Set the type to URL and enter the server URL.

Claude Code. Run this command, restart Claude Code, and sign in through the browser the first time you use it:

claude mcp add fabric --transport http https://mcp.api.fabric.so/mcp
claude mcp add fabric --transport http https://mcp.api.fabric.so/mcp
claude mcp add fabric --transport http https://mcp.api.fabric.so/mcp

Codex. Run this command:

codex mcp add fabric --url https://mcp.api.fabric.so/mcp
codex mcp add fabric --url https://mcp.api.fabric.so/mcp
codex mcp add fabric --url https://mcp.api.fabric.so/mcp

VS Code with Copilot. Add a server named "fabric" to .vscode/mcp.json or to your user settings, with type http and the server URL.

Any other MCP client that supports remote (HTTP) servers works with the same URL, and sign-in is discovered automatically. If sign-in fails, remove the Fabric server from the AI tool and add it again. Items an agent creates count against your Fabric storage and AI usage, which you can check in your Fabric settings.

MCP is the route that makes one drive work across devices. Claude on your phone, ChatGPT in a browser and a coding agent in a terminal can all connect to the same server and see the same files. Your agents, across all your devices shows what that looks like day to day. For a wider look at MCP servers for documents, see best MCP server for documents.


The CLI for terminals and scripts

Some agents work best in a shell. A coding agent in a sandbox, a cron job or a build step often has no desktop app and no chat interface, only a terminal. For these, Fabric has a command-line tool. Through it, agents can run the same server-side search as MCP, with grep, regular expressions and full-text search across the whole drive, and get results back without fetching the files first.

It suits coding agents that already work in commands, and the download page has installation instructions and the current commands.


The API for your own applications

When you're building something rather than chatting with an assistant, use the Fabric API. It's a REST API that covers almost everything Fabric can do: creating and editing notes, managing files and folders, applying and searching tags, creating tasks, adding and managing memories, and searching programmatically. It authenticates with an API key from your workspace settings, and team members can each have their own key. Documentation is at developers.fabric.so.

Content created through the API is searchable and available to Fabric's AI assistant like anything else in the drive. The API is on the Pro and Team plans.

Say your team has a small service that groups each morning's support tickets by theme. It can write the summary into a shared folder through the API, where people read it and Claude or ChatGPT, connected through MCP, can find it later.


Choosing a route

Start from where the agent runs. If it runs on a computer where you can install the desktop app, the mounted drive is the least setup and the most familiar to the agent. If it runs inside an app you don't control, such as Claude, ChatGPT or Cursor, use MCP. If it runs in a terminal or sandbox without a desktop app, use the CLI. If you're writing the software yourself, use the API.

Then think about search. An agent that mostly opens a few known files is fine on any route. One that needs to hunt across thousands of files for a phrase or pattern should use MCP or the CLI, where the search runs on the server.

A small agency might settle on this mix: designers' local agents work in client folders on the mounted drive, account managers use ChatGPT and Claude through MCP to look up briefs, and one script files weekly reports through the API.


The same rules apply on every route

How an agent connects doesn't change what it's allowed to do. In Fabric, access for external agents is set with Agent Vault, which treats them as non-human members of the workspace. You choose the folder, space or project an agent can reach, which actions it can take (reading files, creating documents, moving content, deleting items, running searches), and how long the access lasts, from a single task to indefinitely. Why AI agent access should expire covers why short-lived access is the safer default.

The drive's own safeguards apply too. If an agent saves a change based on an older version of a file, Fabric refuses the write, and the agent re-reads the file and reapplies its change, as when two AI agents edit the same file describes. Agents can read part of a large file without fetching all of it. Instruction files such as CLAUDE.md, AGENTS.md and SKILL.md can't be changed by agents, even ones with write access. Deleted files go to a trash you can restore from, and earlier versions are kept in version history.

Fabric prices per person, so connecting more agents doesn't add seats. Plans are on the team pricing page. For the reasons a team would want all of this in the first place, read your AI agents need a shared drive, not a group chat.


Frequently asked questions

How do I connect Claude to a shared drive?

Add Fabric from Claude's connector directory, or in Claude Desktop open Settings, go to Connectors, add a custom connector with the URL https://mcp.api.fabric.so/mcp, and sign in. Claude can then search and read your Fabric files.

How do I connect ChatGPT to my files?

In ChatGPT, open Settings, go to Connectors, choose Add Connector, enter Fabric's MCP server URL and save. In a new conversation, click More, choose Add Connectors, select Fabric and sign in.

What is an MCP server?

It's a service that follows the Model Context Protocol, an open standard that lets AI tools call outside services as tools. Fabric's MCP server lets an AI tool search, read and create content in your Fabric drive.

How do I add an MCP server to Claude Code or Codex?

For Claude Code, run claude mcp add fabric --transport http https://mcp.api.fabric.so/mcp, restart, and sign in through the browser. For Codex, run codex mcp add fabric --url https://mcp.api.fabric.so/mcp.

Can a coding agent use a mounted drive like a normal folder?

Yes. The Fabric desktop app mounts Fabric as a drive on Mac, Windows and Linux, and agents on that computer read and write it like any other folder. Files stream when opened and take no disk space.

Should an agent use MCP or the mounted drive?

Use the mounted drive for agents running on a computer with the desktop app. Use MCP for assistants without a filesystem, and for agents that need to search a large drive, since MCP search runs on the server.

Do I need an API key to connect an AI agent to Fabric?

Not for MCP, which signs in with OAuth. The API uses an API key from your workspace settings and is on the Pro and Team plans.

Can I control what a connected agent can access?

Yes. Agent Vault lets you limit each agent to a folder, space or project, choose which actions it can take, and set how long access lasts. Every agent action is logged.

Does each connected agent need its own Fabric seat?

No. Fabric prices per person, so agents don't need seats of their own.


Related pages




The workspace that thinks with you.

Ready when you are.

The workspace that thinks with you.

Ready when you are.

The workspace that thinks with you.

Ready when you are.