Skip to main content

Tool Customization for AI Agents

Virtual MCP servers can expose hundreds of tools to AI agents. Without careful curation, this creates problems: degraded AI performance, security risks, and poor user outcomes. This guide explains how to select and manage tools effectively.

The Token Problem​

Every tool you enable adds tokens to the AI agent's context. Consider what happens when an agent sees 200 available tools:

  1. Context consumption: Tool descriptions alone might use 50,000+ tokens
  2. Decision paralysis: The agent struggles to select the right tool
  3. Increased latency: More tools mean longer processing times
  4. Higher costs: Token usage directly impacts API costs / usage limits

The solution isn't to minimize tools blindly but to curate thoughtfully.

Understanding Tool Categories​

Virtual MCP servers classify tools into two fundamental categories:

MCP Tool Categories

Read-Only Tools​

These tools retrieve information without side effects. Examples:

  • Database queries
  • File reading operations
  • API status checks
  • Search functions

Read-only tools are generally safe to enable broadly. The main consideration is relevance: will agents actually use them effectively?

Modifying Tools​

These tools change state or data. Examples:

  • Database inserts/updates
  • File creation/deletion
  • API calls that trigger actions
  • Configuration changes

Modifying tools require careful consideration. Each one represents a potential risk if misused by an agent or compromised account.

For efficiency, use bulk controls to, e.g., enable all read-only tools or disable all modifying tools at once.

  • Enable all read-only tools at once
  • Disable all tools from a specific connector

Selection Strategies​

Begin by understanding what users need agents to accomplish:

  1. User needs: What tasks do they want AI assistance with?
  2. Analyze workflows: Which tools support those specific tasks?
  3. Identify patterns: What combinations of tools work together?

For example, a support team needs agents to:

  • Search documentation (read-only)
  • Query customer data (read-only)
  • Update ticket status (modifying)
  • Never delete customer records (exclude)

This analysis leads to a focused tool set of perhaps 10-15 tools instead of 100+.

The Customization Interface​

Virtual MCP servers provide granular control over tool exposure:

Individual Tool Management​

Individual Tool Toggles

In a Virtual MCP, each tool can be:

  • Enabled/Disabled: Simple on/off control
  • Renamed: Avoid conflicts with other tool names, steer agent behavior by use-case-specific naming
  • Documented: Add context to help agents use tools correctly in your use-case