What tools can AI models call?
Types of tools
Common tools include web search, weather APIs, calendar management, email sending, database queries, and mathematical calculations. Essentially, any function that can be exposed via an API or local code can be a tool.
Custom tools are also possible, such as internal company APIs, file system operations, or IoT device controls. The limit is often the security and practicality of giving the AI access.
Some platforms offer pre-built tools, like retrieval from a vector database or code execution environments.
- Information retrieval: web search, news, weather.
- Productivity: calendar, email, task management.
- Data: databases, spreadsheets, analytics.
- Utilities: calculators, unit converters, translators.
- Custom: internal APIs, file operations, home automation.
How tools are made available
Tools are typically provided to the model through function calling interfaces. The model doesn't execute the tool itself; it generates a structured request that your application handles.
For MCP, tools are discovered from MCP servers. The AI assistant can list available tools and call them as needed.
The range of tools is expanding as more integrations are developed, but each tool must be explicitly defined and connected.
Common mistakes
- Thinking AI can call any tool without integration; it needs explicit definitions.
- Assuming tools are always safe; they can introduce security risks.
- Believing tools are limited to a fixed set; custom tools are common.