When working with Claude Code, organizing your project structure and making relevant files easily accessible to MCP is crucial. MCP’s filesystem capabilities allow you to define which directories and files your Claude agent can see and interact with. Think of it as setting the boundaries of your AI’s workspace. By carefully curating this, you prevent Claude from getting lost in irrelevant data and ensure it focuses on your development tasks.
A common scenario is having separate directories for source code, tests, and documentation. You can configure MCP to grant access only to the relevant directories for a specific task. For instance, if you’re asking Claude to refactor a piece of code, you might only want it to have access to your src directory. This keeps the context clean and the AI’s output more focused.
To achieve this, you can utilize the mcp.yaml configuration file. Here’s a simplified example of how you might specify allowed directories:
agent:
name: CodeRefactorAgent
model: claude-3-opus-20240229
# Define the filesystem scope for this agent
filesystem:
# Allow read-only access to the src directory
allow:
- path: src
readOnly: true
# Optionally, deny access to specific directories
deny:
- path: node_modules
- path: dist
By default, MCP might have broader filesystem access. Explicitly defining allow and deny rules in your mcp.yaml gives you fine-grained control. This is especially powerful when using MCP with shell tools. You can dynamically adjust these settings before invoking an agent for a specific task, streamlining your workflow and enhancing Claude’s understanding of your project.