What is Tool Calling?
Tool Calling is a capability that allows large language models to request the execution of specific functions or external APIs — giving the AI model a way to take real-world actions beyond generating text, such as searching the web, querying a database, or sending an email.
Tool Calling: Full Explanation
LLMs are text-in, text-out systems. Without tool calling, an AI assistant can describe how to check a flight status but cannot actually check it. Tool calling bridges this gap. You define a set of tools (functions) available to the LLM and describe what each one does. When the LLM determines it needs a tool, it outputs a structured request (tool name + parameters) rather than a plain text response. Your application executes the tool, returns the result, and the LLM incorporates it into its response.
Tool calling was popularised by OpenAI's function calling feature (introduced June 2023) and is now supported by all major LLM providers — Anthropic Claude, Google Gemini, Mistral, and open-source models via inference frameworks. The capability is foundational to building AI agents: without tool calling, an agent can only reason; with tool calling, it can act.
Common tools used in enterprise AI applications include: web search (Tavily, Bing API), database queries (SQL over your data warehouse), CRM reads and writes (Salesforce, HubSpot), calendar management, email sending, code execution (Python sandboxes for data analysis), and internal API calls to your own business systems. The combination of an LLM's reasoning ability with your organisation's toolset creates a powerful automation layer.
Key Facts About Tool Calling
- ✓Tool calling allows LLMs to invoke external functions and APIs — moving from text generation to real-world action.
- ✓Supported by all major LLM providers: OpenAI, Anthropic (Claude), Google (Gemini), Mistral.
- ✓The LLM decides when and how to call a tool based on the user's goal — it does not call tools randomly.
- ✓Foundational capability for AI agents — without tool calling, agents can only reason; with it, they can act.
- ✓Tools are defined as JSON schemas describing function name, description, and required parameters.
- ✓Security consideration: tool permissions should follow minimum-privilege principles — agents should not have more access than needed.
How Tool Calling Works
Tool calling works in three steps. First, you define your tools as structured function schemas (name, description, parameters) and pass them to the LLM in the API request. Second, the LLM generates a special response indicating it wants to call a specific tool with specific parameters — rather than a plain text answer. Third, your application executes the tool, captures the result, and sends it back to the LLM as the next message. The LLM then generates its final response incorporating the tool result.
This back-and-forth can happen multiple times in a single "turn" if the LLM needs to call several tools in sequence (or in parallel, which modern APIs support). For example, a sales intelligence agent might call a web search tool, then a CRM lookup tool, then a news API — all before generating its final output.
LangChain and LangGraph provide abstractions that make tool calling easier to work with — you define Python functions and decorate them as tools, and the framework handles the JSON schema generation and result injection automatically.
Real-World Example: E-commerce & Retail
A D2C fashion brand in India built a customer service agent using Claude with tool calling. The agent is given three tools: look_up_order (queries order database by order ID or email), check_return_policy (retrieves policy text from a knowledge base), and create_return_request (writes to the returns management system). When a customer asks "I want to return my order from last week," the agent calls look_up_order to find the order, checks the return policy to confirm eligibility, and creates the return request — all in a single conversation turn, without a human agent.
Frequently Asked Questions
What is the difference between tool calling and plugins?
Tool calling (or function calling) is a native LLM API capability — the model itself generates structured tool invocation requests that your application handles. Plugins (like ChatGPT plugins) were a product-level abstraction built on top of tool calling for end users. In developer contexts, tool calling / function calling is the correct technical term.
Can an LLM call tools in parallel?
Yes — modern LLM APIs support parallel tool calling, where the model requests multiple tool executions simultaneously rather than sequentially. For example, a research agent can call a web search and a database query at the same time, then synthesise both results. This significantly speeds up multi-tool agentic workflows.
How do you control what tools an AI agent can use?
You control available tools entirely in your application — the LLM can only call tools you explicitly provide in the API request. This is an important security property. For sensitive operations like sending emails, modifying records, or making payments, best practice is to add human-in-the-loop confirmation steps before executing the tool, regardless of what the LLM requests.