What is MCP?
Model Context Protocol (MCP) is an open standard for connecting AI applications to tools and data. Instead of every AI client shipping bespoke integrations, MCP defines a small protocol so any client can talk to any server that speaks it.
The analogy that keeps coming up is USB-C: a single connector that replaces a drawer of adapters. In AI terms, it means the same 'GitHub server' can power Claude Desktop, Cursor, Zed, or a custom in-house agent without rewrites.
Why it matters
Before MCP, giving an assistant real tools was an integration project per client. That kept useful AI stuck in vendor silos. With MCP, publishing a tool once makes it available to a growing ecosystem of clients.
For teams building internal AI, MCP dramatically reduces the cost of exposing an internal knowledge base, ticketing system, or data warehouse to whichever assistant your engineers already use.
How it works
An MCP server exposes tools (functions the model can call), resources (data the model can read), and prompts (parameterized templates). The client discovers these at connection time and negotiates permissions.
Servers can run locally as a subprocess or remotely over HTTP + SSE. Local servers are ideal for filesystem, git, or Docker access; remote servers make sense for shared team resources.
- Tools: read_file, create_pr, run_query — actions the model requests.
- Resources: doc://policies/security-2025 — content the model can pull in.
- Prompts: /review-pr — templates surfaced to the user.
Getting started
The fastest path is to install one of the official reference servers into Claude Desktop or Cursor. The filesystem, GitHub, and Postgres servers all work in minutes.
Once you've felt the difference in an existing client, writing your own server is straightforward: pick a language SDK, expose two or three tools, and configure the client to launch it.
When not to use MCP
MCP is not the right layer for user-facing product features. If you're building an app that calls an LLM, keep your tool calls in your own backend where you can enforce business rules.
MCP shines when the goal is to plug a capability into whichever AI client the user chose — internal power users, developers, and research workflows.