In November 2024, Anthropic quietly open-sourced a specification called the Model Context Protocol. Eighteen months later, MCP has become the de facto standard for how AI applications talk to external tools and data. OpenAI added support. Cursor, Windsurf and Sourcegraph baked it in. There are now over 1,000 open-source MCP servers on GitHub covering everything from Google Drive to Kubernetes. If you build anything with AI in 2026 and you are not using MCP, you are probably rewriting integrations that someone else already solved.
This guide explains what MCP is, why it exploded so fast, how it actually works under the hood, how it differs from function calling, and how to start using it in a business setting. Aimed at CTOs, senior developers and technical founders building AI products or internal tools.
What is MCP: the simple definition
MCP (Model Context Protocol) is an open standard for connecting AI applications to external tools, data sources and systems through a single unified interface. Instead of writing a custom integration between every AI app and every data source, you build an MCP server once and any MCP-compatible AI client can use it.
Anthropic describes MCP as "USB-C for AI". The analogy is exact. Before USB-C you had a different cable for every device. Before MCP you had a different integration pattern for every AI tool. MCP is one port, many cables.
MCP is not a replacement for LLMs, function calling or agents. It is the plumbing that makes those things easier to build, because you stop wiring up each integration from scratch.
Why MCP matters (the problem it solves)
Consider what it took to build an AI assistant in 2023 that could read your Google Drive, update your Jira tickets and run SQL queries against your internal database:
- Custom Google Drive integration with OAuth, file listing, content retrieval.
- Custom Jira integration with API authentication and rate limits.
- Custom database connector with SQL injection protection and schema discovery.
- Wiring each of these into the AI's function-calling schema.
- Maintaining it all when APIs change.
- Repeating it all over again if you wanted to switch from OpenAI to Claude.
Each integration was bespoke. Each AI vendor had its own format. Switching LLMs meant rewriting the integration layer. This was the N×M problem of AI tooling: N AI applications × M data sources = N×M integrations.
MCP collapses this to N+M. Write an MCP server once for your data source. Any MCP client can use it. Write an MCP client once for your AI app. It can use any MCP server. The economics are identical to why USB replaced proprietary ports, and why REST replaced SOAP.
How MCP works: clients, servers, primitives
MCP has three core concepts: the client, the server and three types of primitives that flow between them.
Clients and servers
- MCP client: the AI application. Claude desktop, Claude Code, Cursor, Windsurf, ChatGPT with MCP enabled. The client talks to an LLM and needs tools and data.
- MCP server: a small program that exposes a specific capability. A GitHub server. A Postgres server. A Slack server. Your internal ERP server.
- Transport: the communication layer. Stdio (for local servers), HTTP+SSE (for remote) or Streamable HTTP (the current recommended transport in 2026).
Three primitives
Servers can expose three kinds of capabilities to clients:
- Tools: functions the AI can call. "search-github-issues", "query-database", "send-slack-message". The most commonly used primitive.
- Resources: readable data the AI can consume as context. A file, a database row, the current state of a Jira ticket. The client decides when to include these.
- Prompts: reusable templates that the server offers. "summarise-pr-thread" might be a prompt exposed by a GitHub server that takes a PR URL and returns a pre-formatted summary request.
Typical flow
- User installs an MCP server (e.g., the Postgres one) locally or connects to a remote one.
- User opens Claude desktop, which is an MCP client.
- Claude desktop discovers what tools, resources and prompts the Postgres server offers.
- User asks Claude: "show me the 10 most recent orders over £1000."
- Claude decides to use the Postgres server's `query` tool. It passes a SQL query to the server.
- Server runs the query, returns rows.
- Claude summarises the rows in natural language.
The protocol is based on JSON-RPC 2.0 and the spec is open at spec.modelcontextprotocol.io.
MCP vs function calling vs plugins
Three concepts that get conflated. Here is the clean distinction.
| Concept | Scope | Standardised? | Example |
|---|---|---|---|
| Function calling | Inside one LLM conversation | No (vendor-specific) | OpenAI tools, Anthropic tool_use, Gemini functions |
| MCP | Between AI app and external tools | Yes (open standard) | Any MCP client ↔ any MCP server |
| Plugins (ChatGPT, 2023) | Inside ChatGPT ecosystem | No (vendor-specific) | Deprecated in 2024 |
Function calling still exists inside MCP — when a Claude client decides to invoke a tool exposed by an MCP server, it uses Anthropic's tool-use format internally. MCP just standardises how that tool was discovered and described in the first place.
Who uses MCP in 2026
MCP clients (the AI apps)
- Claude desktop — first MCP client, still the reference implementation.
- Claude Code — Anthropic's agentic CLI, uses MCP heavily for dev tools.
- Cursor — supports MCP servers as of mid-2025.
- Windsurf (Codeium) — MCP-native from launch.
- Sourcegraph Cody — enterprise code search with MCP.
- ChatGPT — OpenAI added MCP support in 2025 for developer tier.
- Continue.dev — open-source code assistant, MCP support.
- Goose (Block) — open-source agent framework with MCP.
Popular MCP servers (the integrations)
- Official Anthropic servers: Google Drive, Slack, GitHub, Postgres, Puppeteer, filesystem, Brave Search, Fetch.
- Community servers: Notion, Linear, Jira, Stripe, BigQuery, Redis, Kubernetes, Airtable, Figma, Zoom, Gmail, and hundreds more.
- Enterprise servers: Salesforce, HubSpot, Snowflake, DataDog, Okta — mostly built by vendors themselves in 2025.
Public registries of MCP servers: Anthropic's official repo and awesome-mcp-servers.
Real business use cases
1. Internal knowledge assistant with RAG
Instead of building a custom RAG stack, you can run an MCP server that indexes your internal documents (Confluence, Google Drive, SharePoint) and connect it to Claude or any MCP-compatible client. Users ask questions; the client fetches relevant chunks through MCP and answers with citations. See our broader Claude AI guide.
2. AI agent that manages tickets
An agent connected to a Linear MCP server and a GitHub MCP server can read bug reports, check related PRs, propose fixes and file follow-up issues — all through one AI client, no custom integration code.
3. Sales AI assistant with CRM access
A Salesforce or HubSpot MCP server exposes tools like "get-account-history", "update-deal-stage", "create-task". A sales rep asks the AI assistant: "summarise the last 5 interactions with Acme Corp and log today's call." The assistant uses MCP tools to pull the history and log the call. No custom integration per rep's preferred AI tool.
4. Developer productivity
Claude Code with MCP connected to your internal monorepo, bug tracker and deploy tooling becomes a usable pair programmer. It can check if a test exists, run it, read build logs and propose fixes in the code.
5. Data analysis without exporting data
Connect a BigQuery or Postgres MCP server to Claude. Ask questions in plain English. The AI writes the SQL, runs it, analyses the results. Data never leaves your warehouse; the AI never sees raw rows unless you allow it.
How to get started with MCP
Option 1: Use existing servers
Install Claude desktop. Pick a server from the official registry. Add it to your Claude desktop config. Restart. That is it. Tools from that server are now available in conversations.
Option 2: Build your own server
Anthropic provides SDKs in TypeScript, Python, Java, Kotlin, C#, Swift and Rust. A minimal server is ~50 lines of code. You define:
- The tools your server exposes (name, description, input schema).
- The resources it offers (URIs and content providers).
- Optional prompts.
The official quickstart guide walks through a Python example in about 30 minutes.
Option 3: Remote MCP servers for teams
For enterprise use you usually want remote MCP servers (served over HTTP+SSE or Streamable HTTP) rather than local stdio. Deploy them behind OAuth. Anthropic added first-class remote MCP support in early 2025, and most major enterprise AI clients now support remote servers.
Limitations and security considerations
MCP is production-ready in 2026 but has caveats:
- Authentication is your problem. The spec standardises the protocol, not the auth layer. Remote servers need OAuth, API keys or similar — your responsibility.
- Prompt injection is real. A malicious MCP server can poison instructions that the AI reads. Run only servers you trust. Treat MCP server outputs like untrusted user input.
- Data exfiltration risk. An AI client that has access to both a read-only server (your data) and a write-capable server (Slack, email) can be tricked into leaking data. Sandbox aggressively.
- Performance. Each MCP tool call adds latency. For tight agent loops this matters. Batching helps.
- Discovery is limited. There is no official global registry yet. Finding the right MCP server still means browsing GitHub.
Security guidance: treat MCP servers as you would treat third-party npm packages or Docker images. Audit before use. Prefer official or reputable sources. Run untrusted servers in sandboxed environments. See our related article on DeepSeek vs ChatGPT for more on AI security tradeoffs.
FAQ
What is MCP (Model Context Protocol)?
MCP is an open standard created by Anthropic in late 2024 that lets AI applications connect to external tools, data sources and systems through a single unified interface. Instead of writing custom integrations for every AI app and every data source, developers build MCP servers once and any MCP-compatible AI client can use them. Think of it as USB-C for AI: one port, many cables.
How does MCP differ from function calling?
Function calling is vendor-specific: OpenAI's function calling, Anthropic's tool use, Google's function calling all have different formats. MCP is a standardised protocol that sits above these. An MCP server works with any MCP client regardless of which LLM underneath. Function calling happens inside one conversation; MCP is the plumbing that makes the tools and data available to function calling in the first place.
Who uses MCP in 2026?
Anthropic's Claude desktop app and Claude Code are the primary MCP clients. Cursor, Windsurf, Sourcegraph Cody and several enterprise AI platforms support MCP as of early 2026. On the server side, there are official MCP servers for Google Drive, Slack, GitHub, Postgres, Puppeteer, filesystem access and hundreds of community-built servers. OpenAI added MCP support to ChatGPT in 2025.
Should my company use MCP?
If you are building internal AI tools or agents, yes, MCP saves you from rewriting integrations every time a new model comes out. If you are just using ChatGPT or Claude as a consumer chat product, MCP is mostly invisible. The biggest winners are developer teams building custom AI assistants, and enterprises wanting one standard way for AI to access their systems.