What is the Filesystem MCP server and is it safe to use at work?

Short answer

The Filesystem MCP server is one of the reference servers of the Model Context Protocol project, published as @modelcontextprotocol/server-filesystem. It lets an AI tool read, write, edit, move and search files, within its active allowlist, which older clients can replace through Roots unless you enable --confine-roots. A --read-only flag removes every write tool. The maintainers call it an educational example, not a production-ready solution.

What it does

The Filesystem server gives an AI app a set of file tools: read_text_file, read_media_file, read_multiple_files, write_file, append_file, edit_file (with a dry-run preview and a git-style diff), create_directory, list_directory, directory_tree, move_file, copy_file, search_files, get_file_info and list_allowed_directories. Files inside the allowed folders can also be read as MCP resources.

It runs locally as a Node.js process (npx @modelcontextprotocol/server-filesystem) or as the Docker image mcp/filesystem. A typical use is Claude Desktop: chat apps cannot open files on your disk by themselves, and this server gives them that ability within limits you set.

The limits you set

Everything depends on how you start the server.

Setting What it does
Directories on the command line The allowlist. On protocol 2026-07-28 it is the whole list; older clients can replace it through Roots unless you add --confine-roots.
--read-only Registers only tools marked readOnlyHint: true. Write tools are not listed, and a call to one fails as an unknown tool. Enforced by the server, whatever the client allows.
--read-only-dir <path> Makes one folder read-only while others stay writable.
--confine-roots Lets a client narrow access with Roots, but never widen it beyond the command-line folders.
Docker ro mount Read-only at the operating-system level, as an extra layer.

Roots, the older way for a client to set the folders at runtime, is deprecated in protocol version 2026-07-28. Pass the folders on the command line.

The protocol itself agrees that roots are not a lock. The specification calls them “informational guidance rather than an access-control mechanism” and says the protocol does not enforce that servers stay within them. So the limit has to be enforced by the server, as the allowlist and read-only options above do, and not by a client that merely announces its roots.

Safe setup, step by step

  1. Give it one working folder, not your home directory. A project folder or a shared “AI inbox” folder is enough.
  2. Start read-only for anyone who only wants the AI to read and summarize documents.
  3. Protect reference material with --read-only-dir: the AI can write in its workspace but not in your contracts or templates.
  4. Use dry run for edits. edit_file with dryRun shows the diff before anything changes.
  5. Mind hard links. The README warns that access checks are path-based: a file that is hard-linked from elsewhere changes at every name when edited.

The common misunderstanding

People treat the Filesystem server as an official product because it lives in the Model Context Protocol organization. The maintainers say the opposite: the reference servers are “intended as reference implementations to demonstrate MCP features and SDK usage”, “not as production-ready solutions”. It is well built and its README is careful about security, but nobody guarantees it for your company’s file share. For production use, the usual advice applies: least privilege, a narrow folder, and review of what the agent writes. See is MCP safe to use in a company.

A second misunderstanding: giving an AI your folders does not give it understanding of them. It can search and open files when it thinks of it. It does not know which version of a document is the current one, which folder is an archive, or which file the team agreed is leading.

What this means for a team

The Filesystem server is personal by nature: it runs on one laptop, with that person’s folders. Two colleagues who each connect their own Claude Desktop to their own folders end up with two AIs that know different things. Sharing files on a network drive does not change that, because what each AI concludes from those files stays in its own chats. For a team-wide setup, compare local vs remote MCP servers and read what is an MCP memory server.

Where Bombos fits

The Filesystem server lets one person’s AI open files; Bombos gives every AI on the team the same memory of what is in them and what was agreed. It is a plugin underneath Claude, Claude Code, ChatGPT, Copilot, Grok, Codex, Cursor and your own agents that draws on what already sits in documents, mail and repos, learns how your team works, and shows with every answer what it rests on. It runs in the cloud in Europe or standalone on your own server, and one person can try it first with their own agents.

Related questions

Which folders can the Filesystem MCP server see?

Start it with --confine-roots and explicitly allowed directories. With older clients, Roots can otherwise replace the command-line allowlist. The list_allowed_directories tool shows the active list. Symlinks outside the active allowlist are refused.

How do I make it read-only?

Start it with --read-only. The server then only registers the tools that read and list, so write tools such as write_file and edit_file do not exist for the client. Use --read-only-dir to protect one folder while others stay writable.

Can I run it in Docker?

Yes, as the image mcp/filesystem. You mount the folders under /projects, and adding ro to a mount makes it read-only at the Docker level too.

Do I need it in Claude Code or Cursor?

Usually not. Coding agents such as Claude Code and Cursor already have their own file tools in the project folder. The Filesystem server is most useful in chat apps such as Claude Desktop that otherwise cannot touch your files.

Sources

  1. modelcontextprotocol/servers: Filesystem MCP Server README
  2. modelcontextprotocol/servers: README (reference servers)
  3. Model Context Protocol specification 2026-07-28: Roots
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.