Claude Code remains a strong choice for software development, but a local large language model (LLM) with filesystem access has become the preferred option for routine file management tasks such as organising folders, renaming files and editing markdown documents.
The change was driven largely by privacy considerations. Claude Code reads files into the conversation and sends them to Anthropic’s servers as part of its cloud-based workflow. While that may not be a concern for source code, the arrangement is less appealing when working with personal documents such as tax records or medical files.
Filesystem MCP adds direct file access
The feature enabling this workflow is the filesystem MCP server, a Node.js package published on npm as @modelcontextprotocol/server-filesystem and maintained by Anthropic’s Model Context Protocol (MCP) team.
Once configured, it allows a local model to interact directly with files inside a specified directory. Available actions include:
- Reading and writing files.
- Editing text with diff previews before changes are committed.
- Moving and renaming files.
- Searching folders.
- Viewing file metadata.
- Listing directory contents.
The server operates within a sandbox, restricting access to the folder specified during setup.
Configuration varies by application. In LM Studio, it requires editing the mcp.json file and restarting the application. Jan includes filesystem support in its default MCP server list, while llama.cpp provides an MCP client through its web interface. Ollama does not support MCP natively and requires a bridge such as MCPHost or ollmcp.
Privacy influenced the move
The ability to keep files on local hardware became the main reason for using a local model more often.
According to the source, Claude Code sends files it reads to Anthropic’s servers as part of the conversation. Depending on account settings, the data is retained for either five years or 30 days if data retention has been opted out of.
The source also references a reported prompt-injection exploit affecting Claude’s Code Interpreter path, while noting that the issue should not be overstated. Even so, it argues that keeping sensitive personal folders on local hardware reduces exposure.
Tool access changed how local models are used
Adding filesystem access transformed the role of local LLMs from chat-based assistants into tools capable of carrying out tasks directly on the computer.
The source reports using Qwen 3.5 9B with 8GB of VRAM for multi-step operations such as analysing folder contents, suggesting categories, creating subfolders and moving files automatically. It also notes improved tool-calling performance in Gemma 4 compared with its predecessor.
For routine administrative work, the source says a local model can perform tasks similar to Claude Code without cloud data transfer, usage costs or rate limits, although hardware remains the main constraint.
Hardware limits remain
Filesystem access extends beyond moving files.
The source says the MCP server can rewrite markdown documents, update YAML frontmatter, edit configuration files and perform line-based changes with previewed differences before applying edits. It can also read PDFs, although extraction quality depends on the model being used.
Reliable multi-step tool calling requires sufficient hardware. The source describes 8GB of VRAM running Qwen 3.5 9B as a practical minimum for dependable chained operations, while systems with around 6GB of VRAM may still handle simpler, single-tool tasks.
Claude Code is still described as the stronger option for software development and more complex agentic coding workflows. However, for everyday computer administration, the addition of filesystem access made a local LLM the preferred tool.
TL;DR:
A local LLM equipped with the filesystem MCP server can organise, edit and manage files directly on a computer. The source says this approach offers greater privacy than cloud-based workflows for routine file management tasks, while Claude Code remains better suited to complex coding.
Key points:
- Filesystem MCP enables local LLMs to read, edit, move and rename files within a selected directory.
- The source cites privacy as the main reason for shifting routine file tasks away from Claude Code.
- LM Studio, Jan and llama.cpp support MCP, while Ollama requires a bridge.
- Qwen 3.5 9B on 8GB of VRAM is reported to handle multi-step tool calling reliably.
- Claude Code remains the preferred option for advanced software development work.
Disclaimer : Opinions and investment insights shared by experts are personal and independent. Management assume no responsibility for investment decisions made based on such views. Investors should seek guidance from qualified financial professionals prior to acting.
