Tool use is an AI capability where the model can call external functions, APIs, and tools to take actions beyond text generation. Instead of just describing what should happen, AI with tool use can actually read files, run code, search the web, query databases, and interact with services — making it genuinely agentic.
Introduced by OpenAI in 2023 with function calling, and rapidly adopted by Anthropic, Google, and other model providers. MCP (Model Context Protocol) standardized how tools connect to AI models.
Tool use transforms AI from a conversationalist into an actor. It's the capability that makes agentic coding possible.
| Without Tool Use | With Tool Use |
|---|---|
| Generates text about code | Reads and writes actual files |
| Describes what to do | Does it |
| Suggests commands | Runs commands |
| Imagines results | Observes real results |
| Tool | What It Does |
|---|---|
| File read/write | Access and modify code files |
| Terminal | Run shell commands |
| Web search | Find current information |
| Database query | Read and write data |
| API calls | Interact with external services |
| Code execution | Run code and observe output |
MCP (Model Context Protocol) standardizes how tools connect to AI. Instead of each AI tool building custom integrations, MCP provides a universal protocol — like USB for AI tools.
Tool use is what makes the difference between AI that talks about code and AI that writes, tests, and deploys code. Every agentic coding feature you use — Cursor's agent mode, Claude's computer use — relies on tool use under the hood.