Blog · Explicatif

Qu'est-ce que le
Model Context Protocol ?


If you've seen Claude do things like search the web, read your files, or log a transaction in your finance tracker — that's MCP at work. Model Context Protocol is the open standard that makes it possible for AI assistants to connect to external tools and services in a structured, secure way.

The problem MCP solves

Large language models are trained on static data. Without external connections, they can't check today's weather, read your latest emails, or write to a database. Developers have long worked around this with ad-hoc integrations — custom function calls, bespoke plugins, one-off APIs. Each integration was different, hard to reuse, and only worked with one model.

MCP standardizes the interface. Instead of building a custom integration for each AI model, you build one MCP server and any MCP-compatible client can use it.

How MCP works

MCP defines a client-server protocol. The MCP client is the AI assistant (like Claude). The MCP server is an external service that exposes tools and resources. The two communicate over a defined protocol — either via standard input/output (for local servers) or HTTP with Server-Sent Events (for remote servers).

Tools

Tools are callable functions exposed by the MCP server. Claude can call them on your behalf. A tool has a name, a description (which Claude reads to understand what it does), and a JSON Schema for its input parameters. For example:

  • add-transaction — logs a new transaction to your finance tracker
  • web_search — searches the web and returns results
  • read_file — reads a file from your local filesystem
  • create_issue — creates a GitHub issue

Resources

Resources are data that the MCP server exposes for Claude to read — like the contents of a file, a database record, or in Kachink's case, an interactive dashboard rendered as an HTML page inside the chat.

Authentication

Remote MCP servers (accessed over HTTP) use OAuth 2.0 for authentication. When you connect Kachink to Claude, you're redirected to Google OAuth to prove your identity. Claude then holds an access token that authorizes it to call Kachink's tools on your behalf.

MCP in practice: example servers

The MCP ecosystem has grown rapidly since Anthropic open-sourced the spec in late 2024. Here are some categories of MCP servers people use today:

  • Filesystem — read, write, and search files on your local machine
  • GitHub — create issues, review PRs, search repositories
  • Databases — query PostgreSQL, SQLite, or other databases
  • Productivity — create calendar events, search emails, manage tasks
  • Finance — log transactions, query spending, manage budgets (Kachink)
  • Web — search the web, fetch pages, extract structured data

How Kachink uses MCP

Kachink is a personal finance tracker built entirely as a remote MCP server. When you add https://kachink.app/mcp to Claude, Claude gains access to these tools:

  • add-transaction — log an expense or income entry
  • get-transactions — retrieve transactions for a given month
  • get-month-data — fetch aggregated monthly data for the dashboard
  • add-recurring — set up a recurring monthly transaction
  • edit-transaction — update an existing transaction
  • delete-transaction — remove a transaction

Claude reads the tool descriptions, understands when to use each one, and calls them based on what you ask in natural language. When you say "log $47 for lunch", Claude calls add-transaction with the right parameters — no form, no UI, no context switching.

Kachink also uses MCP Apps — a feature of the MCP protocol that lets servers return interactive HTML UIs. When you ask Claude to "open the dashboard", Kachink returns a full rendered dashboard as an MCP App that appears inside the chat interface.

Getting started with MCP

To use any MCP server with Claude, open Claude's settings, navigate to Integrations, and add the server URL. For remote servers like Kachink, you'll be prompted to authenticate.

If you want to build an MCP server, Anthropic maintains official SDKs for TypeScript and Python. The TypeScript SDK is published as @modelcontextprotocol/sdk on npm.

Ready to try an MCP server? Ajouter Kachink à Claude and log your first transaction — the whole setup takes under a minute.


FAQ

Questions courantes sur MCP

Qu'est-ce que MCP ?

MCP (Model Context Protocol) est un standard ouvert créé par Anthropic qui permet aux assistants IA comme Claude de se connecter à des outils, API et sources de données externes. Au lieu d'être limité aux données d'entraînement, Claude peut utiliser des serveurs MCP pour effectuer des actions dans le monde réel.

Qui a créé MCP ?

MCP a été créé par Anthropic et rendu open source en novembre 2024. La spécification et les SDK sont disponibles sur github.com/modelcontextprotocol.

Quels clients prennent en charge MCP ?

Claude.ai prend en charge les serveurs MCP distants via la page Intégrations des paramètres. D'autres clients incluent Claude Desktop, Cursor, Windsurf et toute application construite avec le SDK MCP. Les serveurs distants (accessibles via HTTPS) fonctionnent directement avec Claude.ai.

En quoi Kachink est-il différent des autres serveurs MCP ?

Kachink écrit et stocke vos données de manière persistante dans une base de données Cloudflare D1, avec isolation par utilisateur et authentification Google OAuth. Il sert également un tableau de bord interactif sous forme de MCP App. Voir le tutoriel de configuration →


Essayer

Découvrez MCP en action avec Kachink

Ajoutez Kachink comme votre premier serveur MCP et découvrez comment le langage naturel se transforme en appels d'outils structurés — sans que vous n'ayez rien à faire.

https://kachink.app/mcp