Model Context Protocol, the open standard for connecting AI assistants to data sources and external tools, has become indispensable in AI developer circles focused on developing and deploying agentic AI. But other developers and many CIOs are just now hearing of the new open-source protocol that rapidly became the leading favorite as a potential standard.
Some believe MCP will soon become an official standard for connecting AI models to tools, data sources, and websites. Others, however, argue that it’s too early to tell which new protocol will win out. This leaves CIOs scratching their heads and wondering when, where and if they need to squeeze MCP servers into their overstuffed AI toolbox and how much the cost may sting their already overstretched IT budgets. Inevitably, that leads to the question of whether it’s smarter and cheaper to build or to buy.
“The tension lies in whether you have the sustained capacity to keep pace with protocols that are still being debated by their maintainers,” said Rishi Bhargava, co-founder at Descope, a customer and agentic IAM platform. “Are you prepared to build the plane while it’s flying, or would you rather upgrade a finished plane mid-flight?”
Adding pressure to the CIO’s decision is the unnerving fact that enterprise leadership is beginning to push back on new AI investments as too few projects make it to production. Those that do get deployed seem to have little impact on reducing payroll or cutting other costs. Leadership’s queasiness is based on more than a feeling.
An MIT report found that 95% of generative AI pilots in companies are failing to achieve significant business outcomes. A 2024 Boston Consulting Group report is slightly more conservative in its AI failure rate, but still dishearteningly high. It states that “74% of companies have yet to show tangible value from their use of AI” and only “a small minority (4% overall) consistently generate impactful results from AI.”
Slight percentage differences in AI failure rates and the debate over their root causes notwithstanding, consensus is growing that AI is failing hard on its promises. According to SentryTech Solutions’ industry analysis, “More than 80% of organizations aren’t seeing any tangible impact on their bottom line from their AI investments.” Put in more relatable numbers, SentryTech found that “despite the headlines, the hype, and the billions being poured into AI technology, four out of five companies are essentially burning money with little to show for it.”
Some AI developers strongly believe that MCP can help turn AI toward more profitable ends.
“MCP is about providing additional tools to AI-driven workflows,” said Blake Crawford, co-founder and CTO of Fusion Collective, an AI consultancy focused on bias detection and responsible technology implementation. “If your AI system needs to integrate with your Slack subscription, your Google Workspace, or even your AWS account, for example, all of that is via MCP. Think of it as a reasonably lightweight method for allowing existing systems to ‘plug in’ to your AI.”
That being said, now would be the time for CIOs to take a hard look at the value MCP adds to your company’s AI capabilities and decide whether, how, and when to move forward.
Build vs. Buy: Strategic Considerations for MCP Servers
According to Abhishek Jain, director of HRIS at Concentrix, an IT services and IT consulting company, the decision itself boils down to the same considerations CIOs have always had to weigh.
“From a business perspective, the build versus buy decision for MCP servers boils down to strategic priorities and risk appetite,” Jain said. Building MCP servers in-house gives you “complete control,” but buying provides “speed, reliability, and lower operational burden,” he said.
But others think there’s no reason to rush your decision.
Michal Prywata, co-founder of frontier AI developer Vertus, argues that the build vs. buy question for MCP servers misses the real issue.
“Most companies shouldn’t be doing either yet,” he said, explaining that companies should first focus on the specific business goals they are trying to achieve, rather than on which existing applications they think should have AI features added.
“Build when you have an actual AI application that requires custom data integration and you understand exactly what intelligence you’re trying to deploy. If you’re simply connecting ChatGPT to your CRM, you don’t need MCP at all,” Prywata said.
When should you buy MCP servers? According to Prywata, whose previous ventures span MIT-incubated medical robotics, agricultural intelligence systems, and comprehensive space technology infrastructure: “Never, honestly.”
Prywata says the MCP ecosystem is “way too new” and vendor controlled. “You’re essentially betting Anthropic’s architecture will become the standard. It might, but that’s a risky bet when the entire AI landscape is shifting every few months,” he said.
It’s also interesting that, unlike other technology investments, there’s not a huge difference in the complexity between buying and building MCP servers.
“An MCP server is not particularly difficult to build or use, which is one of the reasons it has taken off so quickly,” said Tom Moor, head of engineering at Linear, a project management tool for engineering teams that counts the likes of OpenAI and Perplexity among its customers.
Evolving MCP Ecosystem: Risks, Opportunities and Emerging Solutions
However, Moor says that MCP-as-a-service is “definitely a thing” and points to Merge Model Context Protocol as one of many examples.
“There are a number of API companies that allow you to define your API as a specification, and they are well placed to introduce automatic creation of MCP servers. However, you can’t just map one-to-one to a traditional API usually; there is a bit more art to how and what you expose to the LLM,” Moor added.
The twist here is that you don’t really need to “buy” an MCP server today.
Anthropic and the open-source community already provide many prebuilt MCP servers that cover popular productivity tools (such as Google Drive and Slack), developer platforms (including Git/GitHub, Puppeteer for browser automation, etc.), and databases (like PostgreSQL).
According to Xiangpeng Wan, product lead at NetMind.AI, if a specific system doesn’t have a server yet, a company can easily hire a third party or build one in-house. Since MCP is an open standard, anyone can make a compatible server, which, of course, also leaves room for paid, commercial options. Software vendors, he said, may ship official MCP connectors for their products and offer enterprise support.
“That’s one way to ‘buy’ an MCP integration. As for MCP as a Service, it’s starting to appear, but it’s still relatively early in the market, Wan said. Earlier this year, Cloudflare and others rolled out hosted MCP server options, so developers can deploy to the cloud with one click and let end users grant access via OAuth2. “This turns MCP into a managed platform and reduces the ops burden,” he explained.
However, a significant problem is lurking in the vast differences in quality among pre-built MCP servers — CIOs are well-advised to look carefully at the quality of the MCP server they are getting.
“Organizations that support MCPs seem to be inherently better engineered than those built by individual ‘vibe coders,'” said Joseph Ours, partner and AI solutions director at Centric Consulting and an early contributor to FastMCP, which is now the de facto standard for Python-based MCP servers.
What organizations support MCPs? Many do — and more are adopting them every day.
According to Mohith Shrivastava, principal developer advocate at Salesforce, the company’s own AgentExchange presents MCP servers in an “app store-like” environment. Any company needing to connect its AI agent to a specific service can “simply find and acquire a ready-made MCP server from the marketplace, saving significant development time and effort,” Shrivastava said,
Option 3: The Phased Approach
“It is usually best to build [MCP servers] in-house when compliance, performance tuning, or data sovereignty are key priorities for the business,” said Marcus McGehee, founder at The AI Consulting Lab. “Buying a managed MCP solution is ideal when flexibility, scalability, and predictable operating costs are more important than full customization.”
But as it turns out, there is a third option. It’s a carefully phased approach to help protect your call in the continuous flux of AI and protocol evolutions.
“What we’re actually seeing work in practice is a phased approach: buy to learn, build to differentiate,” said Jesse Flores, founder and CEO of SuperWebPros, a website designer company that builds smart websites and helps its clients prepare for AI agents to collect information and buy products online.
He advised that companies start with commercial MCP servers to establish baseline capabilities and understand actual integration patterns; then selectively build where they’ve identified “genuine competitive moats.”
“The key metric is time-to-validated-learning. If you can prove business value with bought infrastructure in 90 days, you’ve earned the credibility to propose a build strategy with actual usage data behind it,” Flores added.
He’s not the only one to suggest this or a similar tactic.
“Many organizations start by buying and gradually internalize [build] as their AI capabilities mature,” said Jain.
