Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in late 2024, that defines how AI applications discover and use capabilities exposed by external programs called MCP servers. Before MCP, connecting a language model to a database, a search engine, or an internal API meant writing a bespoke integration for every model-tool pair. MCP replaces that N-times-M problem with a single protocol: build one server, and any MCP-compatible client can use it.

Clients include Claude Desktop, Claude Code, Cursor, Windsurf, VS Code's Copilot, and dozens of other applications. Servers range from a few lines of code wrapping a local SQLite file to production endpoints serving thousands of agents.

MCP is a client-server protocol built on JSON-RPC 2.0. When a client connects, the two sides negotiate capabilities, then the client can list and invoke what the server offers. The protocol defines three server-side primitives: tools (functions the model can call, like search_flights or run_query), resources (readable data such as files or database records), and prompts (reusable prompt templates).

It also defines client-side primitives that flow the other way: sampling lets a server request an LLM completion from the client, and roots let the client scope which directories or URIs the server may operate within.

MCP supports two main transports. Local servers run as a subprocess on the user's machine and communicate over stdio — the client launches them with a command like npx or uvx. Remote servers run on the operator's infrastructure and speak Streamable HTTP, so a client anywhere connects with just a URL.

The distinction matters in practice. Local servers can touch the filesystem and private credentials but require per-machine setup. Remote servers need zero installation and can serve many users at once, which is why hosted endpoints have become the dominant pattern for commercial MCP services.

The protocol standardizes discovery and invocation, but it says nothing about how a tool author gets paid when an agent calls their server. Authentication exists (OAuth for remote servers), but there is no native metering or billing primitive.

That gap is where the x402 protocol fits. A monetized MCP server intercepts an unpaid tool call, responds with HTTP 402 and a price, and the agent's wallet pays in USDC — settled on Base in about two seconds — before the handler runs. MCP handles what a tool does; x402 handles what it costs.

Loomal indexes thousands of MCP servers — open-source stdio packages and hosted remote endpoints alike — with live-probed tool lists for each, so you can see exactly what a server exposes before connecting. Maintainers can claim a listing by verifying GitHub ownership, attach a per-call price (minimum $0.01), and start receiving x402 payments from agents without changing how their server speaks MCP.