Back to library

🔌Understand the Model Context Protocol (MCP)

Crack open the Model Context Protocol — the host/client/server shape, the three primitives, and the integration math that makes it worth standardizing. By the end you can sketch the MCP server you'd build to expose your own product to AI clients.

Applied14 drops~2-week path · 5–8 min/daytechnology

Phase 1Why a Protocol, Not Another Plugin

Why every AI app kept rebuilding the same plumbing

4 drops
  1. Every AI app rebuilt the same plumbing

    6 min

    MCP exists because every assistant was reinventing tool integration in slightly incompatible ways.

  2. Host, client, server — three roles, not two

    6 min

    MCP splits the AI app into a host that owns the user, clients that own one connection each, and servers that own one capability each.

  3. Resources, tools, prompts — pick the right primitive

    7 min

    MCP servers don't just expose 'functions' — they expose three different shapes, and choosing the wrong one ships a broken integration.

  4. The protocol is boring on purpose

    6 min

    MCP wins by being deliberately uninteresting — JSON-RPC, capability negotiation, no model-specific assumptions.

Phase 2Tracing One Tool Call End-to-End

Trace one tool call from prompt to response

5 drops
  1. Initialize before you do anything else

    6 min

    Every MCP session starts with a capability handshake — and the rest of the session is shaped by what that handshake agreed to.

  2. List before you call

    6 min

    The host learns what tools exist by asking — and the answer can change mid-session.

  3. tools/call is where the model meets the world

    7 min

    Every model-initiated side effect funnels through one method — tools/call — and that's where the host applies policy.

  4. Resources are pulled, not pushed

    6 min

    MCP resources put the host in charge of what context the model sees — the server doesn't get to spam the prompt.

  5. Prompts are the user's slash commands

    7 min

    MCP prompts are server-defined, user-invoked workflows — the third primitive, and the one most engineers forget exists.

Phase 3Primitives, Auth, and the Safety Boundary

Resources, tools, prompts — and the safety boundary

4 drops
  1. stdio for local, HTTP+SSE for remote

    7 min

    Transport choice is trust-boundary choice — stdio inherits local trust, HTTP+SSE recreates it explicitly.

  2. The MCP server you're running as is the one who has access

    8 min

    MCP punts auth to the deployment because the right answer depends on whose trust you're borrowing — OS, IdP, or wrapped API.

  3. The model can't tell a tool from a trojan

    8 min

    Prompt injection isn't an MCP bug — it's a property of mixing data and instructions in one channel. Defense lives in the host's policy, not the server's escaping.

  4. The same capability can be three different primitives

    8 min

    Primitive choice maps to user moment, not to capability — the same data wants to be tool, resource, and prompt in different contexts.

Phase 4Sketch Your Own MCP Server

Sketch the MCP server for your product

1 drop
  1. Sketch the MCP server for your product

    18 min

    Designing an MCP server is mostly about deciding what NOT to expose — the protocol gives you the shape, your product gives you the boundaries.

Frequently asked questions

What problem does the Model Context Protocol actually solve?
This is covered in the “Understand the Model Context Protocol (MCP)” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
How is MCP different from regular function calling or tool use?
This is covered in the “Understand the Model Context Protocol (MCP)” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
What are the three MCP primitives — resources, tools, and prompts?
This is covered in the “Understand the Model Context Protocol (MCP)” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
Who runs the MCP server — the AI provider or the application?
This is covered in the “Understand the Model Context Protocol (MCP)” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.
Do I need an MCP server if I already have a REST API?
This is covered in the “Understand the Model Context Protocol (MCP)” learning path. Start with daily 5-minute micro-lessons that build from fundamentals to hands-on application.