This post is for product and technology leaders navigating the shift from AI-assisted experiences to AI-driven action. It’s not about writing an MCP server — it’s about how to position, execute, and operationalize MCP quickly and responsibly inside a real organization.

---

Yes, MCP is a verb now! There is a good chance you have heard about Model Context Protocol, aka MCP. If not, here is a good intro to the technology. In general, the situation is this: As Agentic AI becomes advanced, users want their AI tools to take action, instead of just talking to them. We want assistants, not just conversation partners. Here is an example.

Say, you are in Turkey, working with Claude to plan the agenda for a business trip — you expect Claude to also seamlessly book your trip from Ankara to Istanbul (which it can do using the Turkish Airlines MCP server). Yes, scenarios like this have been promised for ages — but now it’s actually within reach. MCP servers are cropping up like mushrooms, and AI tools are starting to natively support them. So, the question is not “to MCP or not to MCP” as it once was, but “how to MCP” your product, in a practical, durable fashion, and in just a matter of weeks.

In terms of “how” to do it, I am not just talking about coding an MCP server — that part is getting easier by the week. I am talking about positioning this to executive stakeholders, deciding on an execution model, marketing it, setting support expectations etc. This is where the complexity lies — all companies want to do AI, and many are tinkering with bits and pieces, but few are able to rally the broader organization behind AI efforts, deliver tangible value, and learn iteratively. That’s what this blog is about.

What qualifies me to talk about this? We tackled this very topic at Sinch, a leader in CPaaS. In addition to thinking about AI holistically, our team rapidly released the Sinch MCP server in weeks. This blog draws on that experience.

🎯 First, position it.

Make MCP a strategic capability, not a tech experiment.

If your argument for building MCP capabilities is, “well, everyone else is doing it”, there is a good chance that won’t cut it with executive stakeholders. However, if you center your case around the merits of the protocol — “MCP is the USB-C for AI” — it may garner some interest, but it still may not be strategically deep. Here are some ideas for how to effectively position your efforts.

1. How does MCP server help your customers? I always suggest starting with the customer problem. Any B2C or B2B product provides some service to customers. If you are Stripe, you provide a ‘payment’ service, if you are Turkish Airlines, you provide a ‘flight booking’ service. You typically have public APIs and/or customer facing applications (web, mobile, kiosk) with which your customers consume those services. Now, AI tools are another up-and-coming interface through which your services can be both discovered & consumed. You don’t want to miss out.

2. Discovered and Consumed by AI. This is a key strategic reason why you need to look at MCP. When was the last time in the last few months you searched on Google and clicked on a link? If you search for this very question on Google, there is an AI overview that says 60% of users now rely on the zero-click experience. And all of us regularly use AI assistants — like ChatGPT, Gemini, Claude — for information and guidance. Many of us also use low-code, no-code AI tools like Lovable and Bolt, and every Developer tool is now getting AI powered. What does this mean to you? It means people will increasingly look for your product in AI watering holes. The usual SEO, content, social media marketing efforts alone are not going to cut it — you also need to be discoverable by the AI tools. But wait, AI is not just for ‘discovery’, they also ‘consume’ your services. AI native products like resend.com are seeing tangible uptick in usage from AI building tools. Basically, users need to be able to find your product AND use it through an AI interface, which is where MCP comes in. It is time you explore this topic deeply, or risk being left behind.

3. Provide a harmonized, unified front. This is a secondary argument. In many large companies, each product is specialized, and consequently, it takes some effort to make them work together. Say, the customer uses your Contacts API to retrieve a list of contacts in JSON format, they may need to add a ‘consent’ column before sending it to the Campaigns API to schedule the campaign. With AI in the mix, these ‘integrations’ become easier and cheaper — AI can fill in the gaps. If you provide detailed prompt templates for your MCP server tools — with clear arguments, input validation and output formats — AI will smartly pick the tools, and make them work together. For instance, AI can add the ‘extra column’ seamlessly if you use well defined MCP servers. Also, if you choose to have one MCP server for your suite of products, it projects a unified front to your customers and stakeholders.

🧭 Second, have an execution plan ready.

Move from interest to action with a deliberate execution model.

I suggest being ready with the execution plan as you position MCP with your stakeholders. Most often stakeholders are interested in an idea, but they are not ready to make a go-forward decision because the next steps are not clear. In the fast moving world of AI, such delays are crippling. Here are some ideas for putting together an execution plan

a. Responsible investment with a Preview. Let’s face it — although there is huge promise, MCP servers and toolsets are still evolving. So, I recommend that you designate your MCP server as a ‘preview’ capability. This sends the right signal to your customers and internal stakeholders that “yes, we are not waiting on the sidelines”, but we are also “not jumping in with a big investment yet”. Customers can try it, and give you feedback — and the product team can iterate much more quickly, without committing to solid SLAs, which will come later.

b. Learning by doing. You cannot outsource this effort fully — because having an MCP server is just a small step in your Agentic AI journey. Much of the value is in learning by doing, and getting comfortable with the rapid pace of change. It is like getting familiar with swimming in the rapids. When we started this effort at Sinch, MCP servers were locally deployed. But in weeks, Cloudflare made it easy to host MCP servers, and Claude Integrations capability was available on their $20 per month Pro version. As always, trying out emerging technology has multiple benefits — you can separate the hype from reality as it applies to your product, you can see how customer behavior is evolving, and you can identify some of the strategic changes that are needed in your product.

c. You can execute with a Tiger Team. In a largish company, it may be difficult to find dedicated people to build the MCP server preview. And you may not have the luxury of a Lab or an Alpha division within which to fund this. Enter the Tiger Team. At Sinch, multiple teams were using AI, but they were all busy with their individual roadmaps. So, we came up with an MCP server Tiger team. It was funded by different products, so the learnings can go back to each team. It was cross-functional, so our Marketing or Support teams could learn as well. It had a set timeline — in a few weeks the job will be done, and the team will disband. It had very clear goals — publish a preview of a MCP server to accelerate learning. And it had a single threaded owner. We executed sharply — daily standups, hands-on approach, weekly demos — the whole works!

🚀📢 Third, execute fast, and keep communicating

Unblock delivery by addressing real-world constraints early and often.

Building a basic MCP server itself is not time consuming — you will find 5-minute and 10-minute tutorials. The blockers are the usual suspects — security, testing, documentation etc. Here is how you can tackle those challenges before they block you.

I. Security. How does someone authenticate / authorize with your MCP server? For a local MCP server, the user needs to copy/paste API keys, with the right permissions. For remote MCP servers — the story is more involved. The base MCP protocol itself now works with OAuth 2.1 flows, this is good. But the specification doesn’t go far enough — as per the current version, the MCP server would need to deal with storage for token handling, validating 3rd party tokens etc. Providers like Cloudflare are helping make this simpler, but that is another vendor agreement to get approved. And you may need to coordinate with your Platform / Security teams. Long story short, you can start with a local MCP server as your first goal, as you learn/iterate to release a secure remote MCP server.

II. Issues with your APIs, Apps. When you start exposing your services as an MCP server, you are getting AI tools to work with your APIs. You may find this non-trivial. For example, your API may have a “contact” parameter, and the AI tool may think a phone number or email should work. However, your API may only accept a contact GUID — which it assumes will be available to the caller, who had previously created the contact.. This may have never been a problem for users of the existing APIs. You will discover many such situations, and you will need to work with individual product teams to address them.

III. Testing end to end. You will discover new tools like the MCP inspector for testing and debugging. In some cases, a scenario may work with the MCP inspector, but not on another MCP client — due to differences in transport protocol (SSE vs STDIO) or trust boundaries. Unlike a regular API call or user interfaces — the interactions with AI are much more non-deterministic. The choice of model, the question, the context, the MCP client — each one of them makes a difference. And there aren’t tried-and-tested testing tools available yet, so I suggest having a tight set of scenarios, clear AI tools you are targeting, and getting coverage through a mix of manual and automated tests..

IV. Plan for documentation from the start. AI is advancing at such a rapid pace that the scope of your MCP server needs to be very clear. Which tools has it been tested with? Which LLM models? Which scenarios does it work best in? I suggest starting a rough documentation from the first week and keeping it updated as the work evolves. In our experience, we found that the documentation helped immensely with our testing and communication.

V. Spread the word. Both internally and externally. On the internal side, you will typically find that other people in your company are already working on similar AI efforts, but there was no broader initiative on which they could connect with you. You can use this opportunity to bring them into a fold, guide them, and learn from them without them being threatened. On the external side, the market is hungry for deep, tangible AI efforts. If positioned right, this effort could help boost market perception of your product. At Sinch, we partnered closely with Product Marketing, and highlighted our MCP server effort in many places, including this press release.

Conclusion

As per Gartner, 77% of chief executives think AI is transformative for business, and they are figuring out how to harness it. MCPing your product could be a way to tangibly jump in, try things, and learn. At Sinch, we were able to release an MCP server quickly — in a matter of weeks — with cross-functional participation. You can probably do it faster with the latest advancements. MCPing your product won’t solve everything — but it’s one of the most concrete ways to start learning how AI will actually consume your services. The sooner you start, the faster you’ll separate signal from noise.