Model Context Protocol Changes AI Integration
Anthropic’s Model Context Protocol (MCP) is an open standard to connect AI models to data and tools.
Model Context Protocol
Last week, Anthropic announced Model Context Protocol (MCP), an open standard for connecting AI models to data sources and tools.
Most useful AI applications require connect AI models with data and tools so they can reliably do useful thinking about the world. To fill that need, developers have been able to using various AI frameworks such as Llama Index or LangChain to connect AI models to tools.
So what is new about MCP? Anthropic explains that MCP adds standardization – a single, standard protocol to connect AI models to diverse sources:
[Currently] Every new data source requires its own custom implementation, making truly connected systems difficult to scale.
MCP addresses this challenge. It provides a universal, open standard for connecting AI systems with data sources, replacing fragmented integrations with a single protocol. The result is a simpler, more reliable way to give AI systems access to the data they need.
Standards for Building AI Infrastructure
Providing standardized integration support with MCP is an especially important addition to making AI applications more useful. To explain the need for a standard connector for AI model-tool integration, let’s compare this to the internet.
The internet was built on standards. Standards like HTML, XML, DNS, JSON, the http protocol, etc. all of this standardized plumbing of the internet made it possible to connect from your browser to any website, and then later full-fledged browser applications, on the web. These standards arose and evolved based on the need for common protocols to connect various disparate data and applications.
In the AI era, conventions based on dominant use, like OpenAI API calls on an LLM, have become standard.
OpenAI led with initial efforts to give AI models the ability to execute code, connect to data, and call on tools in 2023. They introduced code interpreter in ChatGPT to run code, function-calling to call tools, and structured outputs (JSON schema outputs) to connect AI outputs to bigger flows. All these are closed solutions to give ChatGPT more capabilities, neither open nor standard.
Early frameworks such as LangChain set the pace for conventions for connecting LLMs and other AI models within flows, providing an open-source library for writing AI applications. As an open framework with robust features, LangChain has became a de facto standard for integrating AI models into flows. However, those integrations are customized for each use case and often bespoke.
MCP’s client-server architecture approach makes it easier to scale connectors to more tools, addressing the scaling issue by developing a standard method of connecting LLMs to resources.
How Model Context Protocol Works
The Model Context Protocol (MCP) is based on a client-server architecture that enables communication between LLM applications, including Claude Desktop and other applications set up to be MCP Hosts, which include MCP Clients that maintain 1-to-1 connections to MCP Servers. The MCP Servers respond to client requests with context, tools, and prompts.
MCP Clients and Servers
MCP is a new path to easier and standardized development of more powerfully integrated AI applications, but it requires a bit of overhead to set up, requiring compliant connectors on both the client and the server side.
The MCP Servers standardize for the LLM how various sources of data or tools interface with the MCP Client. Their job is to respond to tool or data requests from the MCP Client with the appropriate response.
The MCP Client sit within a host application that runs the LLM and the user interface, setting up tool requests depending on what the LLM decides is required for a given user request.
For developers, Anthropic has provided SDKs in Typescript and Python for developers to create MCP servers. The MCP guide shows that developers can set up MCP servers fairly easily. For developers who have built webapps in either Python or Typescript and work with JSON-based, the setup is familiar, and the servers are a few hundred lines of code.
How to use MCP
For Claude users who are not developers, the primary way right now to experience MCP through the Claude desktop app for MacOS or Windows, which you can download here. The Claude desktop app acts as an MCP host and has an MCP client built-in, supporting connections to local MCP servers.
To get started with this, you need to have local MCP servers serving information. Anthropic has helpfully developed and provided several MCP servers in an open-source repository, and partners are adding more. It includes these useful MCP servers:
Web scraping and browser integrations with Brave Search, Puppeteer, and Fetch.
Google integrations with Google Drive and Google Maps.
Database integrations with PostgreSQL and Sqlite.
Utility integrations such as a Filesystem to connect the LLM with local files and data.
The MCP Quickstart guide uses the example of a Sqlite database. One point they emphasize is that the MCP server defines the approved database operations, which gives you a secure interface that maintains control over what the LLM can access.
AI integrations with MCP
MCP will only succeed as a standard if it is adopted in AI integrations beyond Anthropic. Efforts are underway to do that. For example, Anthropic has mentioned several AI coding tool companies have already developed MCP integrations:
Early adopters like Block and Apollo have integrated MCP into their systems, while development tools companies including Zed, Replit, Codeium, and Sourcegraph are working with MCP to enhance their platforms.
The Model Context Protocol GitHub repository links to a number of additional resources and community contributions, which are growing daily.
One contribution is an MCP server connecting to a managed index on LlamaCloud, which in turn can connect to any of many data sources Llama Index supports. MCP competes with Llama Index since both connect LLMs to data. This implementation is an encouraging sign that these frameworks could co-exist and leverage eachother.
MCP clients have been developed that connect AI models beyond Claude, for example, one that connects to Amazon Bedrock API models.
I’d like my local Ollama models to have the power of MCP integrations. Helpfully, Chris Hay presented an MCP client on YouTube that supports multiple LLM sources, including OpenAI and Ollama models such as gpt-4o-mini and qwen2.5.
Challenges and Opportunities for MCP
MCP arose out of the need to standardize and scale connectivity between different data sources and LLMs. While existing AI integrations and frameworks share some similar features, MCP has unique advantages as an open-source framework with a client-server architecture. This is the right paradigm for general inter-connectivity and favors the success of MCP.
That’s the promise of MCP, but the reality is that Anthropic has delivered a beta product. Like their Computer Use release, this implementation is an early partial product that needs to improve top be production-level. They will need the open-source community to adopt and develop it further for MCP to succeed.
Some issues: Debugging can be challenging with complex async client-server stdio communications. MCP addresses plumbing issues but doesn’t guarantee the LLM can make proper use of external information. MCP is not a magic bullet; connectors still need to be developed, tested, and integrated.
MCP could fizzle out if its architecture or implementation ends up inferior to sticking with bespoke approaches that already exist. However, I don’t expect that to happen. MCP’s technical hurdles can and should be overcome. The community adoption and MCP integrations already underway suggest MCP has legs.
We need a standard way to connect AI models to data and tools, and we need that standard to be open and as easy to implement as possible. MCP fills that need with an open standard and a client-server architecture that’s good for interoperability.
MCP has changed how we integrate AI and is on track to be a critical part of AI infrastructure and frameworks.