LLM (Large Language Model)

A large language model is a neural network trained on enormous text corpora to predict what comes next in a sequence. From that single objective, surprisingly general capabilities emerge: answering questions, writing code, summarizing documents, translating, and following multi-step instructions.

Modern LLMs are further trained to follow instructions and to use tools — to recognize when a task needs external data or actions and to emit a structured request for them instead of guessing.

On its own, an LLM can only produce text from what's in its context window; it cannot browse, query a database, or send an email. Tool calling closes that gap. The application presents the model with a list of available tools and their input schemas, and when the model decides a tool is needed it outputs the tool's name and arguments as structured data.

The application executes the call, feeds the result back into the context, and the model continues reasoning with real data. The model decides; the surrounding software acts.

In the Model Context Protocol stack, the LLM is the brain inside the MCP client. The client (Claude Desktop, Cursor, an agent framework) connects to MCP servers, gathers their tool definitions, and exposes them to the model. The LLM then drives the session: it chooses which MCP tools to call, with what arguments, and how to weave the results into its answer.

Quality matters at every layer. A capable model with poorly described tools will misuse them, and a well-built MCP server with vague descriptions will get called wrong — tool descriptions are effectively prompts written for the LLM.

Once an LLM can select and invoke tools autonomously, the natural next question is how those tool calls get paid for. An LLM-driven agent can hit thousands of endpoints in a session — far too many, and far too fast, for human-managed API keys and monthly invoices.

This is where per-call payment rails come in: with x402, the agent's wallet settles a USDC payment on Base before a paid tool's handler runs, so the model's decision to call a tool and the payment for that call happen in one machine-speed exchange. The LLM doesn't manage money itself — the wallet layer does — but its tool choices are what drive the spending.

An LLM is a model; an AI agent is a system. The agent wraps the model with a loop, tools, memory, and goals, letting it take actions over multiple steps rather than produce a single response. Every modern agent contains an LLM, but an LLM alone — with no tools and no loop — is just a very good text generator.