The MCP Layered Model
MCP organizes agent tooling into three capability types: Resources for file-like data, Tools for functions the model can call, and Prompts for pre-written templates. Every skill you publish on ToolVerse plugs into one or more of these layers.
Transports matter too: stdio keeps things local and fast, while streamable HTTP lets remote agents reach your server across the network. Choose stdio for single-host setups and HTTP when skills must be shared across teams.
Pro Tip: Start With One Tool
Resist the urge to expose ten tools on day one. Start with a single, well-tested tool and add capabilities only after real agent traffic validates the shape.
Where Skills Fit
A skill is the thin logic layer above an MCP connection. The server handles transport and permissioning; the skill encodes when to act, how to act, and what to say. That separation keeps infrastructure reusable and behavior auditable.
Transport-Agnostic
Write your skill once and run it over stdio or HTTP without changing business logic.
Auditable Behavior
Because skills live in versioned manifests, every behavior change is reviewable and reversible.
Registering Your First Tool
server.tool(
'get_forecast',
{
description: 'Fetch a weather forecast for a location',
inputSchema: { latitude: 'number', longitude: 'number' },
},
async ({ latitude, longitude }) => {
const url = `https://api.weather.gov/points/${latitude},${longitude}`;
const res = await fetch(url);
return formatForecast(await res.json());
},
);