is
All posts
MCPMCPAnthropictooling

What is an MCP server? A plain-English explainer for builders

An MCP server is a small JSON-RPC service that exposes tools, prompts, and resources to large language models — the wire-format equivalent of a function-calling SDK, but standardised across every model and every client.

inite3 min read

An MCP (Model Context Protocol) server is a small HTTP service that exposes tools, prompts, and resources to large language models over a standard JSON-RPC wire format. The model — Claude Desktop, Claude Code, the Anthropic SDK, or any other MCP-aware client — calls those tools when it needs facts the LLM cannot generate from its weights alone: live database rows, third-party APIs, deterministic math, persistent memory.

Key facts

  • MCP shipped to general availability in Claude Desktop in November 2024.
  • Pre-built MCP servers exist for filesystem, GitHub, Slack, Postgres, and 50+ other surfaces (modelcontextprotocol.io directory, April 2026).
  • ideaudit's MCP server exposes 42 deterministic tools across raw data, math, persistence, and methodology v2.
  • A typical 'audit this idea' loop fires 8-12 MCP tool calls before Claude renders the verdict.

If a function call is the what, MCP is the how — a standard wire format so any tool you build is callable from any MCP-aware client without bespoke glue.

Why MCP showed up at the right time

Three forces converged:

  1. Tool fragmentation collapsed. Before MCP, every model and SDK had its own function-calling shape. MCP gives you one server that works with Claude Desktop, Claude Code, the Anthropic SDK, ChatGPT-with-MCP, Goose, Continue, and everything else that ships an MCP client.
  2. Local-first data wins. MCP servers can run on the user's machine and talk to local files, local databases, and local processes — none of that data needs to leave the user's network for the model to reason about it.
  3. Composition. The same MCP server can be plugged into a desktop chat, a CI pipeline, and an autonomous agent. Tools written once, used everywhere.

What is actually in an MCP server

Three primitives:

  • Tools — typed function definitions the model can call. Each carries a JSON Schema for inputs.
  • Resources — read-only blobs the model can pull into context (files, query results, screenshots).
  • Prompts — pre-written prompt templates the model can ask the user to run.

A vanilla REST API gives you only the equivalent of tools, and you have to teach the model how to call each endpoint by hand. MCP standardises the conversation.

How ideaudit uses MCP

ideaudit is built around an MCP server with 42 tools across four categories:

  • Raw data: Google SERP, Trends, Reddit, Product Hunt, GitHub, Trustpilot, G2, Crunchbase signals.
  • Math: TAM bounds, dealbreaker scoring, social pain index, founder/idea fit, personal-fit-to-investor.
  • Persistence: audits, lenses, entities, mentions, outcomes, collections, investor lists.
  • Methodology v2: cross-lens consensus check, stage-aware verdict synthesis.

When you tell Claude "audit this idea: an AI note-taking app for lawyers", the installed skills tell Claude which tools to call and in what order. Claude does the reasoning on your Anthropic tokens; the MCP server returns deterministic raw data — no hallucinated CAGR numbers, no fabricated competitor lists.

When should you write your own MCP server

Build one when you have:

  • A proprietary data source you want a model to reason over (your CRM, your Notion, your warehouse).
  • Deterministic math the model is bad at (financial projections, SQL, scoring functions).
  • A workflow that needs to be repeatable across sessions and clients.

Skip it when:

  • You're prototyping and a single function call in your SDK app is enough.
  • The data is already public and the model can search for it.
  • You don't need persistence between conversations.

Where to start

The official Anthropic docs at modelcontextprotocol.io have working server examples in TypeScript, Python, and Rust. The community ships pre-built servers for filesystem, GitHub, Slack, Postgres, and dozens more — clone, configure, run.

If you want to use an MCP server before building one, start with ideaudit:

curl -fsSL https://inite.studio/install.sh | sh

That gives you 15 markdown skills + a working MCP integration in your Claude config in under two minutes.

FAQ

Frequently asked questions

  1. How is MCP different from a REST API?

    A REST API is a contract for humans (or programs written by humans). MCP is a contract for models. The protocol is JSON-RPC over stdio or HTTP-SSE, but the higher-level concepts — tools (typed functions), resources (read-only blobs), prompts (templates) — are designed for an LLM to discover and use without bespoke glue.

  2. Do I need to write my own MCP server?

    Not always. Pre-built servers exist for common surfaces (filesystem, GitHub, Slack, Postgres). Build your own when you have proprietary data the model needs to reason over, deterministic math the model is bad at, or a workflow that must be repeatable across sessions and clients.

  3. Which clients support MCP?

    Claude Desktop, Claude Code (the CLI), the Anthropic SDK, ChatGPT-with-MCP, Goose, Continue, Zed, and Cursor — among others. The same MCP server works with all of them; you only need to register it once per client.

  4. How does ideaudit use MCP?

    ideaudit ships 15 markdown skills and an MCP server with 42 tools. When you tell Claude 'audit this idea: …', the skills tell Claude which tools to call and in what order. Claude does the reasoning on your Anthropic tokens; the MCP server returns deterministic raw data — no hallucinated CAGR numbers, no fabricated competitor lists.

Read next