The hands

Tools are actions the harness can take, such as searching the web, reading a file, running code or calling a company system. The model can only ask for them; the harness does the work.

When the model needs something it can't produce from memory, it writes a structured request: "call search_files with Q3 budget." The harness (or, for some tools, the provider's servers) runs it and adds the result to the context as more text. The model reads the result and either asks for another tool or writes its answer.¹² Anthropic's documentation puts it plainly: "The model never runs anything on its own."¹ Some tools run on your side (files, terminal, company systems); some run on the provider's servers (web search, web fetch, code execution).¹⁵ MCP, the Model Context Protocol, is an open standard for connecting AI apps to tools and data.³