AI

What Is an MCP Server?

An MCP server exposes tools, resources, and prompts to AI clients over the Model Context Protocol. Learn what servers do, how they differ from APIs, and examples.

mcp servermcpai agentstoolsintegration

An MCP server is a program that exposes tools, data, and prompts to AI clients over the Model Context Protocol. It is the half of an MCP integration that does the work: the AI client asks, the server acts. A server might wrap a code repository, a database, a SaaS product, or — in Conbersa's case — a fleet of physical phones.

What Does an MCP Server Expose?

MCP servers can publish three kinds of capability:

  • Tools — functions with side effects, such as posting content, sending a message, or starting a job.
  • Resources — readable data, such as files, records, or documents.
  • Prompts — reusable templates and workflows the model can invoke.

The MCP documentation notes that the vast majority of servers today expose tools, which are essentially remote procedure calls. More advanced servers add resources and prompts for richer, two-way interactions (Cloudflare).

How Does a Server Fit Into the System?

MCP uses three roles. The host is the AI application a person uses. The client lives inside the host and translates requests into protocol messages. The server provides the capabilities. IBM's explainer describes the client as the session manager — it handles interruptions, timeouts, reconnections, and error handling — while the server converts requests into actions (IBM).

A single host can run several clients, but each client talks to exactly one server.

How Is an MCP Server Different From an API?

An API is designed for developers who write code against it. An MCP server is designed for a model that decides, at runtime, which tool to call. Practically, a server often wraps APIs and exposes a curated subset of them as tools.

The key difference is intent. A REST endpoint waits to be called correctly by an application. An MCP tool describes itself in a way the model can reason about, so the model can choose it from a menu. IBM stresses that MCP "does not decide when a tool is called and for what purpose" — that remains the model's job.

How Do Local and Remote MCP Servers Differ?

Where a server runs changes how it is used:

Local Remote
Transport stdio HTTP / Streamable HTTP
Users Single user Multiple users
Auth Manual, often env vars OAuth or API keys
Reach Desktop tools Web, mobile, cloud clients

The shift toward remote servers mirrors broader enterprise adoption. Gartner projects that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024, which is why so many servers now ship as authenticated remote endpoints rather than local-only tools (Gartner).

What Are Examples of MCP Servers?

Reference servers exist for Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer, among others — Anthropic shipped several when it launched the protocol. Cursor's documentation points to an ecosystem of official and community servers for tools like Linear, Figma, and browsers.

Increasingly, companies run remote, authenticated servers for their own products. Conbersa is one example: its server exposes a managed device fleet so an agent can post, engage, and schedule (MCP server overview).

The server ecosystem is maturing with the protocol. Postman's 2025 State of the API surveyed more than 5,700 developers, architects, and executives and reported that MCP is early but gaining momentum as a way to expose tools to non-human callers (Postman).

What Should You Check Before Connecting to One?

Security first. OWASP flags "Excessive Agency" — tools and permissions beyond what a task needs — as a top risk when giving agents access to external systems (OWASP). Cursor's own guidance is to verify the source, review permissions, limit API keys to the minimum, and prefer servers hosted by the service provider itself.

Connect to servers you trust, and treat an API key or OAuth grant to an MCP server as real access — because it is. See MCP API key authentication for how to scope it.

How Conbersa Built Its MCP Server

Conbersa runs a remote MCP server over Streamable HTTP at https://api.conbersa.ai/mcp. It exposes read, run, content, and schedule tools that operate real physical devices and accounts. Every request authenticates with a per-user API key, and ownership rules apply per tool, so an agent only sees the accounts its key is allowed to touch. Start with the MCP quickstart.

Neil Ruaro
Founder, Conbersa

We run agentic distribution on a fleet of real phones — and write up what we learn helping founders escape the cold start. Got a topic you want covered? Tell us.

FAQ

Frequently asked questions

An MCP server is a program that exposes external capabilities to AI clients over the Model Context Protocol. It publishes tools the model can call, data it can read, and prompt templates it can use. Examples include servers for GitHub, Slack, databases, and device fleets.
The server provides capabilities; the client consumes them. An MCP client lives inside an AI application like Claude Desktop or Cursor, converts user requests into protocol messages, and calls the server. One host can run multiple clients, each connected to one server.
Both are possible. Local servers usually communicate over stdio and suit single-user tools. Remote servers run over HTTP, serve multiple users, and need real authentication such as OAuth or API keys. Remote servers are what let web and mobile AI clients connect.
The Conbersa Blog

New guides, straight to your inbox.

Tactics on organic distribution and the cold-start problem. What's actually working, no fluff.