MCP Server
Every agent framework used to invent its own way of describing a tool, so an integration written for one agent was useless to the next. An MCP server is the fix: a small program that exposes one system's capabilities in the standard, published shape defined by the Model Context Protocol, so any compliant agent can use it with no bespoke integration code.
What a server exposes
A server speaks JSON-RPC over a transport — stdio for a local process, HTTP for a remote one — and advertises three kinds of primitive:
- Tools: callable functions with schema-typed arguments that the model may choose to invoke — create an issue, run a query, send a message;
- Resources: readable data addressed by URI, pulled into context by the host application;
- Prompts: prepared templates a person can select for common requests against that system.
Everything behind the boundary stays with the server: the API keys, the rate limits, the mapping from a clean tool name onto a messy internal call. Servers exist for filesystems, Git, Postgres, chat platforms and browser automation, and writing one is mostly a matter of declaring your functions with an SDK.
Why it matters
It makes tool use composable. One server, written once, serves a desktop assistant, an IDE extension and your own agent script equally, so the integration problem stops multiplying. It is also the right place to draw the trust line: the server is the only component holding credentials, so narrowing what it exposes narrows what any agent connected to it can ever do.
An MCP server is a program that exposes one system's capabilities to agents in the standard form defined by the Model Context Protocol — tools to call, resources to read, prompts to select — over stdio or HTTP. Written once, it works with any compliant client, removing per-agent integration code. It also holds the credentials for the system behind it, which makes it the natural boundary for access control.
Before standard plugs, every appliance came with its own adapter. An MCP server is the standard plug for AI assistants.
It is a small piece of software that sits in front of something useful — your files, a calendar, a company database — and describes, in an agreed format, what can be done with it: here is what you can look up, here is what you can change, here are the usual requests.
Any assistant that understands the format can use it immediately. Build one server for your system and every compatible AI tool can work with it, instead of each one needing its own custom connection built by hand.