Agents and protocols

Model Context Protocol (MCP)

One standard plug between AI apps and the tools and data they use, and exactly what travels over it.

Intermediate · 7 steps

Step 1 of 7· Illustrative calendar server, tools and data; methods, field names and flow follow MCP revision 2026-07-28

use the steps or arrow keys

Loading visual…

in one minute

Every AI app that wanted to use your calendar, files or CRM used to need its own custom connector for each one. The Model Context Protocol is an open standard, introduced by Anthropic in 2024, that replaces those one-off connectors with one shared language. A service is wrapped once as an MCP server, which describes the tools, data and prompt templates it offers. Any AI app that speaks MCP can then discover and use them. The app stays in charge: it decides what the model sees, sends every call through the protocol and should ask you before anything acts on your behalf. Build once, use in every compatible app.

Why it matters for your product

MCP turns integrations into reusable parts: wrap a system once as a server and it works in any compatible assistant, IDE or agent, which cuts build and maintenance work. It also gives you one place to govern access. Expose only the tools a use case needs, put remote servers behind OAuth with narrow scopes, and make write actions pause for a person. Treat third-party servers like any dependency: their tool descriptions and results are untrusted input that reaches your model.

For engineersShow the maths

integrations to build: custom = N × M; with MCP = N + M

N apps each need a connector to M services when every pair is custom work. With a shared protocol, each app implements a client once and each service ships one server.

Worked example: 4 apps and 5 services: 20 custom connectors, or 9 implementations with MCP. At 10 apps and 50 services, 500 becomes 60.

a cached list is fresh while now < t_received + ttlMs

Since revision 2026-07-28, results of server/discover, the list requests and resources/read carry ttlMs, a freshness hint in milliseconds, so clients can reuse them instead of asking again.

Worked example: A tools/list result received at 10:00:00 with ttlMs = 300000 can be reused until 10:05:00, five minutes later.

where it stops working

MCP standardises how tools are described and called, not whether they are safe or whether the model uses them well: a server's descriptions and results are untrusted text that can carry injected instructions. The protocol is young and still changing; revision 2026-07-28 removed the initialize handshake and deprecated sampling and roots, so clients and servers need compatibility fallbacks.

Key terms