Yesterday we hosted the second episode in our AI webinar series where AJ and I dived deeper into the rapidly changing space of MCPs. If you are not familiar with what MCP is, check out this great overview . In a nutshell, MCP is a protocol introduced by Anthropic which allows you to ‘plug in’ your service into any supported AI tool such as Cursor , Cline , Claude , etc. For data engineers and AI app developers, MCP offers huge potential to not only streamline your development processes, but make it much easier to provide LLMs and agents with customer context.
As we have been building MCP services and working customers, two common approaches to MCP have emerged; the MCP is either context-aware or code-aware.
types of MCP services Context-aware MCP The MCP provides customer/business data to an LLM to offer additional context for an agent to perform analysis. A great example of this is the Shopify Storefront MCP . As a customer, can search products and inventory and add it to your favorite AI tool such as Claude Desktop and do all of your shopping there. What’s more, you can chain the results to ask Claude follow up questions:
“If I buy this jacket from Shopify store x, how does it perform in windy conditions above 5000f?”
Code-aware MCP The MCP is used by developers to improve productivity, build apps, and access your service through their chosen IDE such as Cursor or Cline. There are hundreds of code-aware MCPs available; from GitHub, AWS, Google, and so many more. As a developer, code-aware MCPs eliminate a lot of the work when dealing with a custom API or library dependency. And, with model providers such as openAI, Anthropic, and Google doubling down on MCP support directly within their API calls, code-aware MCPs are a vital tool in the modern developer toolbelt.
Get Started with Airbyte MCP During the webinar, we introduces a number of Airbyte MCP services for both building data pipelines (code-aware) and creating AI apps with customer data (context-aware). As the number of these services grows, an MCP Gateway becomes the cleaner way to expose them all to your agents through one managed connection rather than a tangle of separate servers.
Here’s a quick rundown. You can check out live demos in the webinar recording above.
Embedded-MCP The Embedded-MCP utilizes the new Proxy Request APIs to fetch customer data and files, which can then be used within an AI app or Agent. For example, let’s say you are building an app that uses the OpenAI Responses API , you can use the Embedded-MCP directly within a responses.create request:
PyAirbyte-MCP PyAirbyte MCP is a code-aware MCP designed for data engineers and AI app devs looking to build pipelines as code with PyAirbyte and Airbyte OSS. We’ve created tools for you to list and describe connectors, validate connector config, cache data locally, and stream results directly from sources to LLMs and agents. PyAirbyte-MCP differs from Embedded-MCP in that it uses your data, not that of your customers, and is designed to be lightweight and easily integrated with any Python app.
To get started with PyAirbyte-MCP, make sure you install the SDK on your local machine, or if you have uv installed, and create a file ~/.mcp/airbyte_mcp.env with the following keys, using the GSM standards described here for storing connector credentials.
Then, add the following config to your mcp.json:
Or, if you are using Claude Code, you can add it via:
fast-pyairbyte fast-pyairbyte is another code-aware MCP that allows you to generate an entire data pipeline from a single prompt. All you need to do is specify the source and destination connector you wish to use and let AI do the coding for you. To get started with fast-pyairbyte, check out the recent blog post .
fast-pyairbyte generating a data pipeline within Cursor
What’s next? We are heads down working on AI updates. Check out the change log for regular product announcements. And, make sure you tap subscribe on YouTube to get a reminder of the next episode in the AI webinar series, currently scheduled for September 2nd.
