What Is MCP (Model Context Protocol)? Plain-English Guide (2026)
MCP (Model Context Protocol) is an open standard created by Anthropic that lets AI models call external tools, read data sources, and use prompt templates through a single, consistent interface. It works like a universal adapter: any MCP-compatible AI client can connect to any MCP server without custom integration code. The protocol handles discovery, authentication, input validation, and structured output.
The formal definition
Model Context Protocol (MCP) is a client-server protocol that standardizes how large language models (LLMs) interact with external capabilities. An MCP server exposes tools, resources, and prompts. An MCP client (like Claude Desktop) discovers and calls those capabilities on behalf of the user. The protocol uses JSON-RPC 2.0 over either stdio (local processes) or streamable-http (remote servers) transport.
Why MCP exists
Before MCP, every AI tool integration was bespoke. If you wanted Claude to check your analytics, you wrote a plugin. If you wanted it to schedule a post, you wrote a different plugin. Each integration had its own auth flow, its own data format, and its own error handling. Nothing was reusable.
MCP standardizes this. A tool developer writes one server that works everywhere. An AI client implements the protocol once and gains access to every server. The result is an ecosystem where tools compose freely, similar to how USB standardized hardware peripherals.
How MCP works: the four-step flow
When the client starts, it reads your config file and connects to each MCP server listed there. It asks each server "what tools, resources, and prompts do you offer?" The server responds with a manifest listing everything available, including input schemas, descriptions, and annotations.
When you ask Claude a question, the model sees the tool manifest and decides which tools (if any) are relevant. For example, "score this hook" triggers score_hook. "Analyze @creator" triggers analyze_account. The model picks based on the tool description and your prompt.
The client sends a JSON-RPC request to the server with the tool name and parameters. The server executes the function (calling APIs, querying databases, running analysis) and returns structured results. All of this happens over the transport layer, either stdio or streamable-http.
The model receives the structured output and incorporates it into its response to you. If the tool returns a recommended_chain (like Hooklayer does), the agent can review and execute follow-up tool calls, creating a multi-step workflow from a single prompt.
Stdio vs streamable-http transport
MCP supports two ways for the client and server to communicate. The choice depends on where the server runs.
- Server runs as a local process on your machine
- Client starts the process and pipes JSON-RPC through stdin/stdout
- Config uses
"command"and"args"fields - Requires npm/node installed locally
- Best for: development, private tools, air-gapped environments
- Server runs on a remote host (cloud, SaaS)
- Client connects via HTTPS to the server URL
- Config uses
"url"and"headers"fields - Nothing to install locally
- Best for: SaaS tools, shared teams, production use
Hooklayer uses streamable-http transport. You paste one config block with your API key, restart Claude Desktop, and the tools are live. No npm install, no local processes.
The three MCP primitives: tools, resources, prompts
Tools
Tools are functions the AI agent can call. Each tool has a name, description, input schema (what parameters it accepts), and output schema (what it returns). Examples: score_hook accepts hook text and returns a 0-100 score. analyze_account accepts a TikTok handle and returns viral DNA analysis. Tools also carry annotations like readOnlyHint and destructiveHint that tell the agent whether the operation is safe to auto-execute.
Resources
Resources are data endpoints the agent can read. Think of them as files, database rows, or live data feeds. A CRM server might expose customer records as resources. A file system server might expose project documents. Resources are read-only and use URI-style addressing.
Prompts
Prompts are reusable instruction templates a server can offer. A marketing server might provide a "brand audit" prompt that structures the conversation for competitive analysis. Prompts are less common than tools but useful for complex, multi-step interactions that benefit from predefined structure.
Who uses MCP?
MCP adoption spans several categories of users and organizations.
Content marketers use MCP servers like Hooklayer, the QA gate and slop filter for AI-generated content, to score hooks, analyze competitors, and remix viral content directly from Claude Desktop. No code required. The agent handles tool selection and chaining.
Developers use MCP to give their custom agents access to external APIs without writing wrapper code for each integration. The protocol handles discovery, auth, and structured I/O.
Software companies are publishing MCP servers alongside their REST APIs to make their tools accessible to AI agents. This is becoming a distribution channel: appearing in the agent's tool list means getting used when the agent decides it is the right tool.
n8n, Make, and Zapier have added MCP client nodes. This lets no-code teams build workflows where an AI agent calls MCP tools as part of a larger automation pipeline.
Tool annotations: how agents decide what is safe
Every MCP tool can declare metadata annotations that guide agent behavior. These are especially important for marketing tools because some operations (like scheduling a post) have real-world consequences, while others (like scoring a hook) are read-only.
readOnlyHint: trueTool only reads data, never modifies anything. Safe to auto-execute.destructiveHint: falseTool does not delete or permanently alter data.idempotentHint: trueCalling the tool twice with the same input produces the same result.openWorldHint: trueTool interacts with external services beyond the local environment.All 8 Hooklayer tools declare readOnlyHint: true and destructiveHint: false. They analyze and score content but never modify external accounts or publish anything.
Frequently asked questions
What does MCP stand for?
MCP stands for Model Context Protocol. It is an open standard created by Anthropic that defines how AI models communicate with external tools, data sources, and services through a consistent interface.
Is MCP only for Claude?
No. MCP is an open protocol that any AI model or client can implement. Claude Desktop, Cursor, Windsurf, n8n, and many other clients support MCP. Any developer can build an MCP server that works with all compatible clients.
What is the difference between stdio and streamable-http transport?
Stdio transport runs a local process on your machine. The client starts the process and communicates through standard input and output. Streamable-http transport connects to a remote server over HTTPS. Hosted MCP servers like Hooklayer use streamable-http so you do not need to install anything locally.
What are MCP tools vs resources vs prompts?
Tools are functions the agent can call (like score_hook or analyze_account). Resources are data the agent can read (like a file or database). Prompts are reusable instruction templates that configure how the agent interacts with a server. Most marketing MCP servers primarily expose tools.
How do I install an MCP server?
For hosted servers (streamable-http), paste a JSON config block into your claude_desktop_config.json file and restart Claude Desktop. For local servers (stdio), you install the package via npm and reference the command in your config. Hosted servers are simpler because there is nothing to install.
Is MCP secure?
MCP supports OAuth 2.1 with PKCE for authentication. API keys are stored locally in your config file, never sent to the AI model. Each tool declares security annotations (readOnlyHint, destructiveHint) so agents know which operations are safe to auto-execute. HTTPS encrypts all streamable-http traffic.
Who created MCP?
Anthropic created and open-sourced MCP in late 2024. The protocol specification is publicly available and maintained by the community. Anthropic also runs the Official MCP Registry where verified servers are listed.
What is the Official MCP Registry?
The Official MCP Registry is a directory maintained by Anthropic listing verified MCP servers. Servers published there go through a review process. Hooklayer is listed as io.github.khan-ashifur/hooklayer v1.0.0. The registry also feeds into aggregator sites like PulseMCP.