What is an MCP server?
An MCP server is a program that exposes tools, resources or prompts to AI applications over the Model Context Protocol, so any compatible client can use a system without a custom connector.
Checked
In the Model Context Protocol, the host is the AI application, a client inside it holds a connection, and the server provides capabilities. Common servers wrap a file system, a database, GitHub, Slack or a calendar. The server is written once, and every compatible host can use it.
This page covers the server side: what a server exposes, how it is reached, and what changed for server authors in the 2026-07-28 specification. The protocol itself, its history and its governance are covered on the Model Context Protocol page.
What a server exposes
Tools, controlled by the model
Schema-defined functions the model decides to call, listed through tools/list and run through tools/call. They can write to databases, call APIs or change files.
Resources, controlled by the application
Read-only data with a URI, such as file contents or a database schema, that the host retrieves and decides how to pass to the model.
Prompts, controlled by the user
Parameterised templates a user invokes explicitly, often as slash commands.
Local and remote servers
The specification defines two standard transports. With stdio, the client launches the server as a subprocess and they exchange newline-delimited JSON-RPC messages over standard input and output, which suits local tools. With Streamable HTTP, each message is an HTTP POST to a single endpoint, and replies come back as JSON or a request-scoped event stream, which suits shared remote servers.
What the 2026-07-28 specification changed for servers
Stateless core
The initialize handshake and the Mcp-Session-Id header are gone. Each request carries its protocol version and client capabilities in _meta, so any request can reach any server instance behind a plain load balancer.
State through handles
A server that needs state returns an explicit handle from a tool, and the model passes it back as an argument.
Routable and cacheable
Streamable HTTP requests carry Mcp-Method and Mcp-Name headers for gateways, and list responses can include cache lifetimes.
Deprecations
Roots, Sampling and Logging are deprecated with at least twelve months of support, as is the older HTTP+SSE transport. Client ID Metadata Documents replace Dynamic Client Registration as the preferred registration route.
Extensions
Long-running Tasks moved into an opt-in extension, alongside MCP Apps and Enterprise Managed Authorization.
Where MCP servers go wrong
Too many tools
Every tool definition costs context. Anthropic's tool-writing guidance favours a few tools that cover whole workflows over one wrapper per API endpoint.
Untrusted descriptions
The specification says tool descriptions and annotations should be treated as untrusted unless the server is trusted, and hosts must get user consent before invoking a tool.
Credentials near the model
Tokens held inside the agent's environment can leak through prompt injection. Anthropic's Managed Agents keep MCP OAuth tokens in a vault and call tools through a proxy.
How Motiontic's engineers apply it
Motiontic's engineers build an MCP server when a client system will be used by more than one agent or application, keep its tool set small and its output terse, and scope its credentials to the narrowest permissions the tasks need. Third-party servers are reviewed like any other dependency before an agent can reach them.
The two standard MCP transports
| stdio | Streamable HTTP | |
|---|---|---|
| How it connects | Client launches the server as a subprocess | HTTP POST to one endpoint |
| Typical use | Local tools on the same machine | Shared or hosted servers |
| Replies | Newline-delimited JSON-RPC on standard output | JSON or a request-scoped event stream |
| Scaling | One process per client | Any instance behind a load balancer |
Common questions
What does an MCP server do?
It wraps a system, such as a database or a SaaS API, and exposes its tools, data and prompt templates in the standard MCP format, so any compatible AI application can use them.
What is the difference between an MCP server and an MCP client?
The server provides capabilities. The client lives inside the AI application and connects to servers on the host's behalf, one client per server connection.
Should an MCP server run locally or remotely?
Local stdio servers suit tools tied to one machine, such as a file system. Remote Streamable HTTP servers suit shared systems that many users or agents reach, and need proper authorization.
Are MCP servers safe to install?
Only as safe as their code and permissions. The specification treats tools as arbitrary code execution, so third-party servers deserve the same review as any dependency, with credentials scoped tightly.
Related
Model Context Protocol (MCP)
The open standard for plugging tools and data into AI agents.
Skills vs MCP
Connectivity to systems, compared with packaged expertise the agent loads on demand.
Tool search
Load tools when they are needed, not all at once before the work starts.
Programmatic tool calling
The model writes a script that calls the tools, so raw results never fill its context.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.