All posts
    Blog2 min read

    Picking the Right MCP Protocol

    Sammy Deprez

    When working with the Model Context Protocol (MCP) — the standard that lets AI models talk to tools and data — one of the first choices you’ll face is how the model and tool should communicate.

    There are a few ways to do it: stdio, streaming HTTP, and the older SSE (Server-Sent Events).
    Each has its place, and picking the right one depends on what you’re building.

    Protocols


    1. STDIO — simple and local

    This is the easiest to start with.
    With stdio, the client actually starts the MCP server as a child process — just like running another program from the command line.
    Once it’s running, they communicate through standard input (stdin) and output (stdout), which means the client sends requests as text, and the server replies the same way.

    No network traffic, no ports, no web servers — just two programs talking directly over their input and output streams.
    It’s fast, clean, and great for local tools.

    ✅ Best for: local tools, command-line utilities, or editor extensions.
    ❌ Not for: anything remote, shared, or long-running in the cloud.

    If you’re testing your MCP server or building a local integration (like a VSCode or terminal tool), stdio keeps things simple.
    You can always move to HTTP later when you go remote.


    2. Streaming HTTP — scalable and modern

    This is the newer, more flexible option.
    It runs over regular HTTP but allows streaming, so you can get live updates while the model is still working — instead of waiting for the full result.

    ✅ Best for: web apps, remote MCP servers, or services with multiple clients.
    ❌ Downside: a bit more setup — you’ll need to handle authentication, networking, and possibly scaling.

    If your MCP server will live on the web or handle requests from different users, streaming HTTP is the right tool for the job.


    3. SSE — the older way

    SSE (Server-Sent Events) was used in early MCP specs.
    It let the server push updates to the client, but only in one direction, which made it less flexible.
    Newer versions of MCP replaced this with streaming HTTP, which supports two-way communication.

    ✅ Works fine,
    ❌ But not ideal for new projects.

    Unless you’re maintaining older code, you can safely skip SSE today.


    Quick summary

    Use casePick this
    Local dev toolsstdio
    Cloud or multi-user servicesstreaming HTTP
    Legacy setupsSSE (only if needed)

    Final thoughts

    • Start with stdio for quick local development — it’s fast and zero setup.
    • Move to streaming HTTP when you need scalability or streaming updates.
    • Keep SSE only for old projects or compatibility reasons.

    Choosing the right transport early keeps your MCP setup clean, simple, and future-proof.
    It’s a small detail that saves you from big refactors later on.

    Keep reading.

    New notes, straight to your inbox.

    One email when something new ships. No spam, unsubscribe anytime.

    Cookie Preferences

    We use cookies to enhance your browsing experience. Essential cookies are necessary for the website to function. You can choose to accept all cookies or only essential ones. Learn more in our Privacy Policy