MCP Tool Filtering

Restrict which MCP tools an agent can see and call

Filter MCP tools to control which capabilities agents have access to, improving security and reducing complexity.

from praisonaiagents import Agent, MCP

mcp = MCP("npx -y @modelcontextprotocol/server-filesystem /tmp", allowed_tools=["read_file", "list_directory"])
agent = Agent(name="Reader", instructions="Read files only — never write.", tools=mcp)
agent.start("List files in /tmp and read the smallest one.")

The user restricts which MCP tools the agent sees; only whitelisted tools appear in the tool list.

How It Works

Quick Start

1. Whitelist Specific Tools

   Only allow safe file operations:

   from praisonaiagents import Agent, MCP
   
   agent = Agent(
       name="safe-fs",
       instructions="Read files when asked",
       tools=MCP(
           "npx -y @modelcontextprotocol/server-filesystem /tmp",
           allowed_tools=["read_file", "list_directory"],
       ),
   )
   agent.start("What files are in /tmp?")

   The user asks about files on disk; MCP tools are filtered before the model sees them.

2. Blacklist Dangerous Tools

   Block destructive operations:

   from praisonaiagents import Agent, MCP
   
   agent = Agent(
       name="no-destructive",
       instructions="Help with GitHub tasks safely",
       tools=MCP(
           "npx -y @modelcontextprotocol/server-github",
           disabled_tools=["delete_repository", "delete_issue"],
       ),
   )

How It Works

Precedence Rule: If both allowed_tools and disabled_tools are specified, allowed_tools wins (include takes precedence over exclude).

Configuration Options

Common Patterns

Safety-First Agent

Restrict to read-only operations:

from praisonaiagents import Agent, MCP

# File system - read only
safe_agent = Agent(
    name="reader",
    instructions="Help users understand file contents",
    tools=MCP(
        "npx -y @modelcontextprotocol/server-filesystem /workspace",
        allowed_tools=["read_file", "list_directory"]
    )
)

Remove Dangerous Operations

Block specific risky tools:

from praisonaiagents import Agent, MCP

# GitHub without destructive actions
github_agent = Agent(
    name="github-safe",
    instructions="Help with GitHub tasks safely",
    tools=MCP(
        "npx -y @modelcontextprotocol/server-github", 
        disabled_tools=[
            "delete_repository",
            "delete_issue", 
            "delete_pull_request",
            "force_push"
        ]
    )
)

Multi-Server Filtering

Apply different filters to different servers:

from praisonaiagents import Agent, MCP

# Combine filtered servers
agent = Agent(
    name="secure-developer",
    tools=[
        # Filesystem - read only
        MCP(
            "npx -y @modelcontextprotocol/server-filesystem .",
            allowed_tools=["read_file", "list_directory"]
        ),
        # GitHub - no deletions
        MCP(
            "npx -y @modelcontextprotocol/server-github",
            disabled_tools=["delete_repository", "delete_issue"]
        )
    ]
)

When to Use Which Filter

Use allowed_tools when:

• Security is critical
• You know exactly which tools you need
• Working with untrusted or powerful servers

Use disabled_tools when:

• You want most tools but need to block a few
• Rapid prototyping with tool discovery
• Working with generally safe servers

Best Practices

Related