AI

AI Tool

An external function or service that an AI agent can call autonomously to interact with the real world: APIs, web search, code execution, file reading.

Main robotique articulée en métal brossé tenant une clé plate en verre transparent dont la mâchoire luit d'une pointe bordeaux, posée sur un sol sombre réfléchissant sous un faisceau de lumière blanche
Tool IA — Image générée par IA (Qwen-Image 2.1, ComfyUI local)

An AI tool is an external function or service that an AI agent can call autonomously to interact with the real world. Tools turn a passive LLM into a system able to act: web search, API calls, code execution, database queries, reading and writing files.

Technical definition

Function calling

The core mechanism is function calling. The developer declares the available tools to the LLM as JSON schemas (name, description, expected parameters, types). During inference, the model can decide to generate a structured tool call rather than text. The intermediate system executes the call and returns the result to the model.

Tool categories

Tools fall into families:

  • Search — Web search engines, vector databases (RAG), internal document search
  • Data — SQL queries, REST/GraphQL API calls, reading CSV/JSON files
  • Action — Sending emails, creating tickets, writing to a CRM, posting to a Slack channel
  • Computation — Running Python code, mathematical calculations, data analysis
  • Perception — Image analysis, audio transcription, OCR on scanned documents

Tool vs Skill vs Plugin

These three terms refer to closely related but not identical concepts:

  • Tool: the standard technical term — a function an LLM can call via function calling
  • Skill: a marketing term whose meaning varies — at Microsoft Copilot, it's an application integration; at Claude, it's a complex contextual instruction
  • Plugin: a legacy OpenAI term (deprecated) — replaced by tools in the GPT-4 API

Current challenges

Security and permissions

An agent with access to powerful tools (database writes, sending emails, code execution) needs a strict permission model. The "human in the loop" pattern makes it possible to require human validation before any irreversible action. Without guardrails, an agent can cause real damage.

Description quality

An agent's performance depends directly on the quality of its tool descriptions. A poorly described tool (ambiguous parameters, vague description) will be misused by the LLM. Engineering tool descriptions is a skill in its own right.

Frequently Asked Questions

What is the difference between a tool and a skill?

In most contexts, tool and skill are synonyms: a specific capability an agent can activate. The nuance varies by vendor. At Anthropic (Claude), skills are complex contextual instructions. At Microsoft (Copilot), skills are application integrations. At OpenAI, everything goes through tools (function calling). In practice, both terms refer to the same concept: a function the agent can call.

How does an LLM call a tool?

Through the function calling mechanism. The developer describes the available tools to the LLM (name, description, parameters). When the model judges it relevant to use a tool, it generates a structured call (JSON) with the right parameters instead of generating text. The system executes the call and returns the result to the model, which incorporates it into its answer or decides on a next action.

Resources and Documentation