MCP Server

An MCP server is the piece of software that does the actual work in the Model Context Protocol. It wraps something useful — an API, a database, a browser, the local filesystem — and exposes a defined set of tools that an AI agent can discover and call.

The agent doesn't need to know how the work happens. It sees a tool name, a description, and an input schema; the server handles authentication, API quirks, and data shaping behind that interface. One server, written once, works in every MCP client.

Servers communicate over JSON-RPC and can offer three kinds of capability. Tools are the workhorses: functions with typed inputs the model invokes to act — search the web, run a query, create an issue. Resources expose readable data by URI for loading into context, and prompts are reusable templates that package recommended workflows.

Most servers in the wild are tool-centric. A typical one registers between one and a couple dozen tools, each with a description that doubles as instructions to the calling model.

Distribution falls into two camps. Local servers ship as installable packages — npm (run with npx), PyPI (run with uvx), Docker/OCI images, or one-click MCPB bundles — and run as subprocesses on the user's machine, talking to the client over stdio. Remote servers run as hosted endpoints reachable over Streamable HTTP, with nothing to install.

The trade-off is control versus convenience: local servers can touch local files and keep credentials on-device, while remote servers centralize updates, scale independently, and are the natural shape for anything monetized.

The official MCP registry catalogs published servers from their server.json manifests, and indexes like Loomal build discovery on top — categorized listings, live-probed tool lists, and claim verification so you can tell whether a listing is maintained by the actual author. When evaluating a server, the tool list tells you more than the description: it's the exact surface an agent will see.

MCP standardized how agents call tools but shipped no payment layer, so most servers earn nothing regardless of usage. The x402 protocol fills the gap at the HTTP level: a priced tool call returns 402 with payment requirements, the agent's wallet pays in USDC, settlement lands on Base in about two seconds, and only then does the handler run — with an Ed25519-signed receipt in the response.

On Loomal, a developer who has claimed their server can attach a per-call price starting at $0.01. The platform fee is 5% on settled transactions, currently waived, and there are no chargebacks because on-chain settlement is final.