Glossary · Protocol

What is MCP (Model Context Protocol)?

The Model Context Protocol, or MCP, is an open standard that lets AI assistants connect to external systems — reading their data and calling their functions — through one interface, rather than through a bespoke integration per assistant.

Start freeRead the glossary
[01] THE LONGER ANSWER/ PROTOCOL

Why mcp (model context protocol) matters

Before MCP, connecting an assistant to a product meant building something specific to that assistant, and doing it again for the next one. MCP inverts that: the product exposes a server describing its tools once, and any client that speaks the protocol can use it. Anthropic published it in late 2024 and it has since been adopted well beyond Claude — Cursor, VS Code, ChatGPT and Gemini all connect to third-party MCP servers now, by quite different routes. For a product, shipping an MCP server is the difference between being something people open and being something their assistant can use on their behalf.
[02] HOW IT WORKS/ IN PRACTICE

A server advertises tools

An MCP server publishes a list of tools with names, descriptions and typed arguments. The assistant reads that list and decides which to call — the description is effectively documentation written for a model rather than a person.

Transport is HTTP or stdio

A local server runs as a process and speaks over standard input and output. A remote one is reached over HTTP. Streamable HTTP is the current remote transport; the older HTTP+SSE variant is deprecated, and several clients now only fall back to it.

Authentication is the server’s business

Most remote servers use either an API key in an Authorization header or an OAuth flow. Which one you can use depends on the client — some support headers, some only OAuth, and that decides which setup route works for you.

Scopes still apply

An MCP server is a view onto an API, not a way around its permissions. A key with read-only scopes gives a read-only assistant, which is the sensible default for a client you paste a credential into.
[03] WHAT IT IS NOT/ CONFUSED WITH

MCP (Model Context Protocol) against the terms it gets mixed up with

Most of the confusion in this category is vocabulary rather than substance. These are the terms mcp (model context protocol) is most often used interchangeably with, and the difference that actually matters in each case.

A plugin or an integration

A plugin is usually specific to one host application. MCP is a protocol, so one server works with every client that implements it — which is the entire point and the reason it spread.

Function calling

Function calling is a model capability: the model emits a structured request to run a named function. MCP is the transport and discovery layer around that — how the model learns which functions exist and how the call reaches another system.

An API

An MCP server usually sits in front of an API rather than replacing it. The API is for your code; the MCP server is the same surface described in a way a model can discover and call without you writing the glue.

[04] WHERE IT MISLEADS/ THE CAVEAT

What a definition will not tell you

MCP client configuration is the least stable thing in this space and the most likely thing you will read stale instructions about. The config file key differs between clients — Cursor uses one name and VS Code another — remote entries need an explicit type in one and no type in the other, and the transport flag defaults to the wrong thing in two popular CLIs. Any setup guide older than a quarter is worth checking against the client’s own documentation before you trust it.

What FishOut does about it

FishOut is an MCP server as well as a REST API, on every plan including Free, over Streamable HTTP. Connect a client and it can search your mentions, read relevance, labels and Reply calls, read the Daily brief, add keywords, teach FishOut what is not for you, draft a reply or hand a mention to a teammate — under the scopes of whichever key it authenticates with.

Questions

Read next

The FishOut MCP serverMCP setup for Claude CodeMCP docs

Other terms

Social listeningBuyer intentShare of voiceBrand monitoringSentiment analysisGenerative engine optimisation (GEO)Webhook