What is MCP (Model Context Protocol)?
MCP is Anthropic's open standard for how AI agents talk to external systems. Here's what it is, who supports it in 2026, and whether it's production-ready.
MCP stands for Model Context Protocol. It's an open standard from Anthropic that lets AI models connect to external systems (files, databases, APIs) in a standardized way. For SMBs, that means an AI agent can read your CRM, check your calendar, and update your database without you building a custom integration for each system.
This guide explains what MCP is, how it differs from OpenAI's function calling, which vendors support it in 2026, and whether it's actually production-ready for SMBs right now.
What is MCP really?
MCP is an open protocol (MIT license) that standardizes how AI models fetch data and call tools in external systems. Anthropic launched it on November 25, 2024 to solve a concrete problem: every AI agent build previously required custom integration against every system.
The architecture is simple: an MCP client (your AI app, e.g. Claude Desktop) connects to an MCP server (a process exposing resources from an underlying system). Communication happens via JSON-RPC per the official specification.
Three building blocks in an MCP server
Three primitives define what an MCP server can expose:
- Resources: data the AI model can read: files, database rows, API responses
- Tools: functions the AI model can call: send email, create Git commit, query Postgres
- Prompts: predefined context templates the AI model can use
Everything is open and documented at the modelcontextprotocol GitHub organization, which holds the official specification, the Python SDK, and the TypeScript SDK. The interest is real: the reference-server repo alone had passed 86,000 stars by May 2026, and the specification repo sat at roughly 8,400. Exact figures are perishable and climbing fast. Since 2025 the project has been run under the Linux Foundation, no longer by Anthropic alone.
How does MCP differ from OpenAI Function Calling?
The difference is standardization vs vendor-specific. OpenAI Function Calling is a proprietary feature in OpenAI's API. MCP is an open protocol that any AI model can implement, and any developer can build MCP servers for.
| Aspect | MCP | OpenAI Function Calling |
|---|---|---|
| Standard | Open (MIT license) | Proprietary OpenAI |
| Reusability | One MCP server works with any MCP client | Function definitions written per OpenAI project |
| Discovery | Built-in capability listing | Manual declaration per call |
| Ecosystem | 200+ community-curated servers via awesome-mcp-servers | No shared ecosystem |
| Vendor lock-in | None | Locked to OpenAI |
In practice: if you build against OpenAI and later want to switch to Claude, you have to rewrite all function definitions. With MCP you only swap the client. All MCP servers still work. Same logic against Anthropic, LangChain Tools, or custom REST APIs: no reusability outside their own ecosystem.
Which vendors support MCP in 2026?
As of 2026, MCP is supported by all three major model providers. It's production-ready in the Claude ecosystem, OpenAI adopted it officially in 2025, and Google has built support into Gemini. Anthropic drives the protocol and has the most ready-made servers, but it's no longer Claude-bound.
- Claude Desktop (Anthropic's own app) has full native MCP support since November 2024
- Cursor IDE has native MCP integration per Cursor's own documentation, with support for both MCP tools and resources
- Continue.dev supports MCP natively per their documentation, as do Cline and several other developer tools
- OpenAI officially adopted MCP in March 2025 and shipped full MCP support in ChatGPT in September 2025; the rollout is covered in InfoQ's report
- Google DeepMind joined in April 2025, and support now lives in the Gemini ecosystem
The status shifted during 2025
The picture above moved fast. When this guide was first written, MCP was in practice Claude-bound, but during 2025 the other major players joined. OpenAI integrated MCP broadly across 2025 and now calls it "a key part of how we build", per the protocol's first-anniversary report, and Google and Microsoft have built in support too. The takeaway for you: MCP is no longer a Claude-only protocol but a standard backed by the three largest model providers. That means the reason to choose MCP — avoiding lock-in to a single vendor — has become stronger than it was when the protocol was Claude-bound.
The size of the ecosystem
The official MCP registry started as a grassroots project in February 2025 and launched in preview on September 8, 2025. So the timeline is: the protocol in November 2024, the registry just under a year later. The registry held around 9,600 servers in May 2026, and Anthropic reported more than 10,000 active public servers in December 2025. Among them are official servers from Anthropic for filesystem, Git, Slack, GitHub, Postgres, SQLite, and Google Drive.
Note that the two figures in this guide measure different things: the table above lists 200+ servers in awesome-mcp-servers, a community-curated list where someone manually picked out interesting servers, while the registry counts every published server in the official registry. Those are two data points, not a contradiction.
For a broader introduction to how AI agents work (where MCP is a central protocol), read our guide on what an AI agent is. MCP connects the model to systems; another way to give it knowledge is RAG, which retrieves answers from your own documents.
Which use cases fit SMBs?
Three concrete use cases already deliver value to SMBs: connecting an AI assistant to the business system for real-time lookups, automating reports from internal databases, and letting support fetch customer data without switching tools. They are sorted below from lowest to highest implementation complexity.
Three worked examples by complexity
1. Consulting firm with GitHub + Slack in Claude. A 5-person consulting firm installs MCP servers for GitHub and Slack locally on Claude Desktop. Consultants can ask Claude "summarize the past 24 hours of Slack discussion in the #client-project-x channel and create a GitHub issue for follow-up". Time saved: 10–15 minutes per day per consultant. Risk: low, everything runs locally.
2. E-commerce with Postgres MCP for catalog and inventory. An e-commerce business with 15,000 products exposes the product database via Postgres MCP server. The sales team asks Claude in natural language: "Which products in the electronics category have stock under 10 and haven't sold anything in the past 30 days?". Gets answers with concrete product IDs. Risk: medium, requires DPA if customer data is exposed.
3. Marketing agency with Google Drive + Notion for proposal generation. An agency connects Drive (for existing proposals) and Notion (for client notes) to Claude via MCP. Asks Claude "build a proposal draft for Client X based on the latest meeting notes and our latest similar proposal from Q3". Time saved: 1–2 hours per proposal. Risk: medium-high, requires DPA + clear data access policy.
What an MCP build actually costs with us
The figures above are generic. What we can speak to ourselves is what it costs to go from idea to a deployed MCP integration for a business, since that's work we sell. An MCP build is fundamentally the same thing as a regular AI agent build: a scoped integration against one or more systems. So the price tiers follow our standard implementation levels.
| Implementation level | One-time cost | What it covers with MCP |
|---|---|---|
| Small scoped build | from 10,000 SEK | one local MCP server against a single system, e.g. filesystem or a database |
| Standard build | 25,000–55,000 SEK | two to three servers with authentication, e.g. Postgres and Slack with OAuth |
| Complex build | from 55,000 SEK | many systems, telephony, and ongoing development |
Delivery time runs 2–6 weeks depending on the number of integrations, and maintenance costs 1,500–5,000 SEK per month if you want a contract. If you just want one thing built and then to manage it yourself, you essentially only pay for AI usage afterward. For the full cost picture, including building it yourself and ready-made tools, see our guide on what AI agents cost for businesses.
How does your company get started with MCP?
The easiest way to start with MCP is Claude Desktop + a local MCP server for a specific use case. It requires no developer and takes under 30 minutes to set up. From there you can expand to more servers and heavier integrations once the value is proven.
Concrete steps:
- Download Claude Desktop from claude.ai/download. A Pro subscription is required for MCP usage.
- Choose an official MCP server from github.com/modelcontextprotocol/servers. The filesystem server is easiest to start with.
- Configure the server in Claude Desktop's
claude_desktop_config.jsonper the instructions. - Restart Claude Desktop. The server is now available, and Claude can read and write to chosen directories.
When you need a developer
For scaling, you'll likely need a developer. Postgres MCP, Slack MCP, and Google Drive MCP require API keys, OAuth flows, and security review. In our own builds, a server like that typically takes a couple to a handful of hours to configure in a production-safe way, while the rest of the time goes to testing edge cases and getting the permissions right. That cost lands in the price tiers in the table above. For GDPR aspects and EU AI Act implications, read our guide on the EU AI Act for businesses.
Is MCP a hype technology or production-ready? Honest answer: production-ready, and no longer Claude-bound. During 2025 both OpenAI and Google joined, so MCP is now backed by the three largest model providers. Maturity still varies between clients, so test the specific support in your tool before you build for real, but the lock-in argument against MCP no longer holds. It's precisely the breadth of support that makes the protocol interesting.
Frequently asked questions
No, not for simple use cases. Claude Desktop with an official MCP server for filesystem or GitHub only requires installation and a config file. More advanced integrations like Postgres or custom APIs require developers. In our builds, a server like that typically takes a couple to a handful of hours to configure in a production-safe way.
The protocol itself is free and open. What you pay for is the AI model's usage when it uses MCP. A Claude Pro subscription suffices for individual use. If you want a consultant-built and maintained setup, our maintenance runs 1,500 to 5,000 SEK per month; otherwise you only pay for usage.
Local MCP servers (filesystem, local database, Git) are low risk because data never leaves your machine. Cloud servers (Slack, Google Drive) require DPA agreements with vendors and often DPIA. IMY hasn't given specific guidance on MCP as of May 2026, so treat it like any API integration from a data protection standpoint.
Yes. OpenAI adopted MCP officially during 2025 and shipped full MCP support in ChatGPT in September 2025, and Google has built support into the Gemini ecosystem. So it's no longer a Claude-only protocol. Claude Desktop was first out of the gate, but the three largest model providers now support MCP.
LangChain is a Python/JavaScript library for building AI applications. MCP is a protocol for communication between AI clients and external systems. They don't compete directly. A LangChain app can use MCP servers, and MCP servers can be written in any language including with LangChain libraries.
AI to work?