A developer recently made the switch from Anthropic’s Claude Code to a local large language model (LLM) after realizing the limitations of cloud-based tools. As they put it, ‘For a folder of tax paperwork or medical PDFs, the arrangement gets a little harder to love.’ This experience highlights a growing concern among users about the privacy implications of uploading sensitive documents to external servers.
The solution that caught this developer’s attention is the filesystem Model Context Protocol (MCP) server. This tool transforms local LLMs into powerful file management utilities. It grants models the ability to read, write, and restructure files within a designated directory, all while keeping the rest of the system secure. The server is a Node.js application, and it’s supported by the same team at Anthropic responsible for the MCP framework.
Getting the filesystem MCP server up and running isn’t too complicated. In LM Studio, which is a widely used interface for local LLMs, users need to modify the mcp.json configuration file and restart the application. For Jan, this feature is already included in the default server list. On the other hand, llama.cpp users can activate the functionality in their web UI by using the --webui-mcp-proxy flag. However, Ollama is a bit different. It doesn’t natively support MCP, so a bridge like MCPHost or ollmcp is necessary to integrate it with the filesystem server.
The main reason this developer moved away from cloud-based tools is privacy. When using Claude Code, every file that gets processed is sent to Anthropic’s servers. Depending on the user’s data settings, the files might be stored for up to five years. More recently, there have been reports of prompt-injection attacks affecting the Code Interpreter tool, which has led to increased hesitation among users handling sensitive data.
According to the developer, enabling filesystem access turned their local LLM from a passive chatbot into an active file management assistant. The model itself, the weights, and the VRAM usage all stayed the same. What changed was the access to tools that allowed the model to interact with files directly on the disk. This shift significantly expanded the range of tasks the LLM could perform.
When it comes to model performance, the Qwen 3.5 9B and Gemma 4 models stand out for their ability to handle tool calling tasks effectively. These models can execute complex operations, such as sorting through large numbers of files. While tools like Claude Code still lead in coding-related tasks, the combination of a local model and filesystem access provides a free, private, and efficient way to manage files without the costs or restrictions of cloud services.
Limitations and Practical Use
The developer uses this local LLM setup primarily for administrative work. This includes tasks like sorting files, cleaning up markdown, and managing messy folders. The main constraints are hardware-related, particularly VRAM limitations, which affect the size and complexity of tasks the model can handle. Still, for non-coding tasks, the system works efficiently, eliminating the overhead of cloud-based solutions.
Now, this developer’s workflow is centered on control and privacy. Although they don’t necessarily dismiss cloud-based tools, they find the growing capabilities of local models to be a compelling alternative when it comes to managing sensitive files or handling large-scale file operations.