phimage
← Writing

Has anyone tried MCP? — and what it could mean for 4D

#ai#mcp#4d#agents

Has anyone here actually tried the Model Context Protocol yet? It only shipped in November 2024, so it’s far too young to call an industry standard — but I keep running into people building genuinely useful things on top of it: database access, custom business logic, IDE assistants. Worth a look before it’s everywhere.

What it is, quickly

MCP is a standard way for an AI model to reach external tools, resources, and context through one consistent interface — instead of every app inventing its own glue. The moving parts are roughly: tools (functions the model can call — a DB search, an API call, a computation), resources (documents, tables, knowledge the agent can reference), and context/prompts (session state and reusable templates that keep behavior predictable). The payoff is the usual one for a good standard: add a capability once, and any MCP-speaking client can use it.

The reference implementations are on GitHub, and the “awesome” lists are already filling up fast (punkpeye, appcypher).

Why I’m interested: 4D

I wrote up a (yes, too verbose 😅) exploration doc thinking through what MCP could be for 4D. Two directions stood out.

1. A developer assistant for the 4D IDE. An MCP server that exposes 4D-aware tools to an agent: code generation and completion that understands 4D syntax, methods, and classes; debugging help; database-structure management; form-design support; documentation generation. This is the shape the JetBrains MCP plugin already takes for IntelliJ — read the open file, search the project, run a configuration — so the pattern is proven, just not yet pointed at 4D.

2. Business-logic tools for end apps. The more interesting one for our customers. You register your own 4D operations as tools — CreateInvoice, SearchCustomer, ScheduleDelivery — and expose specific tables or views as resources, with session and permission context around them. A 4D application then becomes a bridge between an AI agent and your business logic, your data, and your external APIs. Users ask in natural language; the agent calls the right registered operation; your code stays in control of what’s allowed.

The bet

If MCP does become the standard it looks like it might, the win is that a 4D app could expose its logic to any MCP-speaking agent without a bespoke integration per tool. That’s a future worth getting ahead of — which is why I’m poking at it now, while it’s still early enough that “has anyone tried this?” is a reasonable question.

Worth reading

Ask me again in a year whether MCP became a standard. My bet is yes — and that the 4D apps which expose their logic through it will look a lot smarter for having started early. 🔌

EDIT — I stopped wondering and started building

A while after writing this, I did roughly what I’d sketched above — in both languages.

  • Swift MCP — I forked the official MCP Swift SDK as phimage/swift-sdk (Apr 2025) to experiment with a child-process transport: spawn a server, talk to it over stdio. That “spawn a process, speak JSON-RPC” shape is the one I’ve kept running into everywhere since.
  • 4D MCP — then the 4D side: an MCP server, mesopelagique/MCP, built on a small JSON-RPC core, mesopelagique/JSONRPC.

So: yes, I tried MCP — in Swift and in 4D — and the bet from this post still looks good.

← Back to all writing