MCP Adoption: what we're learning from early implementers (must read for techies)

In January, I predicted that 1 of the biggest limiting factors to widespread AI integration was going to be APIs…luckily Anthropic had the same thought and launched Model Context Protocol (MCP), a universal plugin for LLMs to easily access and query the underlying data in any application.

It’s already considered table-stakes for developer tool companies.

MCPs are also becoming mainstream. Earlier in June, ChatGPT launched MCPs in their platform through "Connectors" -- meaning you can access your Gmail, HubSpot, SharePoint, Github, etc., directly from within ChatGPT. (Claude has already had this for a month or two). Many developers are already experimenting with it within Cursor.

Last night, I went to a great event with Alex Rattray from Stainless , Dina Kozlov from Cloudflare , Mat Varughese from Stripe , Chris Bell from Knock , and Brooke McKim from Vantage and they discussed their use cases and lessons learned from being early adopters of MCP.

Here are the top use cases and top watch-outs for MCPs - the next wave of tech sweeping the software industry.

Interesting MCP use cases:

• Developer documentation search has emerged as the universal entry point. Most MCP users are accessing MCPs through Cursor, using natural language instead of dealing with building or finding the right API endpoints. This vastly accelerates the building and deployment of products and integrations.
• Stripe allows users to take action in the platform through natural language rather than having to navigate their 1000 API endpoints. They built their MCP servers on top of their existing AI Tool Shed, which theoretically will allow any employees to generate compliant tools.
• At Cloudflare, engineers are using MCP to automatically deploy firewall rules across their entire infrastructure.
• Knock processed 100s of email templates for migration automatically - a task that would have taken weeks before.

MCPs are also enabling additional use cases that transcend APIs.

• Stripe built a "slash me" endpoint specifically to give AI context about the user's access levels, environments, and accounts. They're also implementing auto-tagging functionality and debugging tools.
• At Cloudflare, you can use MCPs to analyze your own behavior and have the LLM suggest additional services or activities you can do within Cloudflare — this could become an upsell functionality.
• Square developed an MCP server that allow you to take a photo of a physical menu and automatically build an online store.

BUT, given the tech is so new, here are 6 watch-outs as you explore MCPs.

Watch-out 1: Business-Specific Definitions

• LLMs are (kinda) bad at math. So if a Stripe user wants to use an LLM to calculate their monthly recurring revenue, it may get it wrong. Stripe is exploring semantic layers that translate natural language queries into platform-specific business logic and enforce Stripe's definitions and calculation methodology.
• Business-specific terms also becomes challenging when a user can access dozens of MCPs through 1 interface. How does the AI know that they want the Stripe MCP and not the QuickBooks MCP, when they are both named Billing (for example)? So companies are all adding their brand name to both MCPs and parameters that may exist in many places.
• Definitions of what data means is also complicating things. You need to be specific enough so an LLM knows what is it looking at, but not so long that you overload the context window. There is a lot of work to be done here on naming and defining your data in LLM-friendly ways, which can get super complex when you think about giant databases.

Watch-out 2: Discoverability

• When you remove the need for a dashboard and people can interact with your product from an AI chat window, how do users know what to ask for? As Cloudflare found with their 15 MCP servers, employees can't find or choose the right tools.
• The best practice that all panelists agreed on is to focus MCP creation on common use cases — you don’t need to cover every API endpoint. But there's still work to be done here to solve discoverability.

Watch-out 3: Security

• Enterprises currently have no way to monitor which MCP servers their employees connect to or what data flows out. Especially as it becomes easier to 1-click and add an MCP to existing software. This creates compliance risks and may slow enterprise rollouts.
• There's emerging discussion about proxy services and middleware layers to manage MCP connections.
• Then there’s the risk of “ignore all previous instructions” prompt injections anytime you connect to the internet through an MCP.
• Each of the companies in the event built their own custom auth servers to control access rights to data (a startup opportunity to standardize this?)
• Then there’s the issue of how powerful LLM inference is. Stripe experimented with deliberately restricted their MCP's ability to read customer data, but the AI still estimated customer numbers by counting invoices - technically accurate but also problematic.

Watch-out 4: Accuracy and Troubleshooting

• The potential for sub-agent architectures - where one MCP leverages another - raises Qs about error handling across multiple systems.
• Companies don’t have access to see what users are querying in their MCP servers. So when things go wrong, clients come to you and troubleshooting gets quite complex.
• Running good Evals before you release your MCP become really important for this reason — and the more complex MCPs become, the harder this will be. (A startup opportunity here?)

Watch-out 5: Cost Concerns

• When the AI makes dozens of API calls to figure something out, those tokens cost money.
• Then layer in an MCP interacting with other MCPs and LLMs making errors and repeating calls, how do you attribute the costs? How do you flow them through to users? How do you prevent costs from ballooning?

Watch-out 6: Scaling Issues

• Scale is revealing critical limitations. You couldn’t expose an entire enterprise to an MCP, because the LLM would run out of context window space to even handle the query.
• Plus LLMs lose the plot when there’s too much information.
• Then there’s tool limits - Cursor restricts users to 40 tools, forcing companies to break down complex operations.
• The result is a move toward use-case-specific MCP servers rather than comprehensive API exposure.
• Stainless is exploring whether LLMs should write code instead of making waterfall natural language calls, which could be more cost-effective.

The promise of accelerating AI through MCPs is proving real, but requires careful navigation of tech constraints and security (and a whole bunch of technical definition writing).

For my technical friends out there, anything particularly interesting or surprising here? What else have you found as you explore MCPs?