Model Context Protocol fever spreads in cloud-native world
The Anthropic-led spec for AI agent tool connections gains further momentum this week, with support from cloud-native infrastructure vendors such as Kubiya and Solo.io.
An emerging standard that connects AI agents with data sources and tools has gained momentum and new features recently, including support from cloud-native infrastructure companies.
AI agents are autonomous software components backed by large language models (LLMs) that can take autonomous action and invoke external tools to accomplish tasks. Model Context Protocol (MCP), introduced by AI vendor Anthropic in November, is one of multiple frameworks for orchestrating and integrating AI agents that have arisen amid last year's expansion of agentic AI.
But MCP has generated growing buzz in the industry, including partnerships with other big generative AI (GenAI) players and hundreds of integrations from IT tools vendors. Microsoft and OpenAI both rolled out new integrations with MCP following its March 26 update: OpenAI launched support for it in its Agents SDK and will soon support it in its ChatGPT desktop app and Responses API, while Microsoft, which already supported MCP in Copilot Studio and extensions for its AutoGen AI agent framework, added an MCP server extension for its Playwright web testing and automation tool set.
This integration caught the eye of one enterprise site reliability engineer (SRE) supporting Amazon Bedrock Agents as part of his company's internal developer platform, where developers also work with Playwright.
"I will have to use it to confirm, but it seems very useful to make it easier for developers to consume [Playwright]," said Mahender Singh, site reliability lead for a financial services company he asked not be named. "We have specialized testers or QA [engineers] doing this [now, but] with this, developers could pick and choose which components to implement in an interactive way. That is powerful."
On Tuesday, AWS jumped aboard the MCP bandwagon with a set of MCP servers for its code assistants. The same day, during KubeCon + CloudNativeCon EU, cloud-native player Kubiya launched an agentic AI platform that runs on Kubernetes and integrates MCP.
"We have our own container orchestration engine … that because of its maturity is even more evolved than MCP and runs stateless tool execution versus needing to work with just servers," wrote Kubiya CEO Amit Govrin in an email to Informa TechTarget this week. "That said, MCP has become a standard protocol, and rather than fighting this trend (not a smart thing for anyone other than a major CSP to do), we are embracing it and plugging into their robust ecosystem where MCP is another 'tool' our users can interface with."
Govrin said MCP support will allow Kubiya users to run the platform directly from where they work, such as Anthropic's Claude desktop, OpenAI, Cursor or any other gateway that supports MCP.
AI agents are more sophisticated than other LLM-driven apps, but they need frameworks like Model Context Protocol to work together.
MCP roadmap and fine print
Developer desktops are primarily where MCP is in use so far, according to industry analysts.
"In theory, an AI agent could be exposed as a 'tool' or set of tools in an MCP server. Another agent could then use it to complete a task, but this isn't what all the buzz is about at the moment," said Gary Olliffe, an analyst at Gartner. "The authentication/authorization model for MCP is limited, and that means it's best for its original intended purpose of allowing AI apps -- a chat service, for example -- to access external tools."
The MCP March 26 update added support for user authentication through Oauth 2.1, but users must still bring their own authorization and access control tools or source them from a vendor. The update also added support for HTTP streaming, which will help support remote MCP server access beyond the desktop.
MCP is still a bit of a science project in many ways, and a lot has to be done to make it work.
Rob StrechayAnalyst, TheCube Research
Still, outstanding items on MCP's roadmap include other features important to remote access, such as service discovery and support for stateless operations required for serverless computing environments.
"When Anthropic supports it and the other [big AI players], you will get momentum," said Rob Strechay, an analyst at TheCube Research. "But they [have] large engineering teams. MCP is still a bit of a science project in many ways, and a lot has to be done to make it work."
Another outstanding roadmap item is a more formal multi-vendor governance structure, which Olliffe said is a crucial next step.
"It would be great for the community if Anthropic clarifies that, in light of the wide adoption and exploration," he said.
Some industry watchers also expressed concern at MCP's launch that Anthropic's proprietary Claude AI model and desktop client are the primary means of deploying MCP, though subsequent releases have added support for other frontier LLMs.
However, none of this has stopped cloud-native infrastructure vendors from joining the MCP ranks and filling some of those gaps themselves. Object storage vendor MinIO previewed its MCP server for its enterprise AIStor product March 28. CDN vendor Cloudflare added its own remote MCP servers accessible from the Internet on March 25. Cloud-native networking vendor Solo.io, which included support for MCP in its kagent project last month, on Wednesday introduced MCP Gateway, an extension of its kgateway open source API gateway project. MCP Gateway provides a centralized tool registry and access point, authentication, authorization and monitoring support.
"We think this is going to be a massive deal," read a Cloudflare blog post. "Connecting coding agents to MCP servers has blown developers' minds over the past few months, and remote MCP servers have the same potential to open up similar new ways of working with LLMs and agents to a much wider audience."
In the meantime, another open source project led by Cisco, AI agent framework maker LangChain and evaluative AI vendor Galileo launched on March 6. The project, called AGNTCY, includes an agent connect protocol (ACP), an agent directory for discovering and publishing agents, a schema framework for a standard agent metadata format and verification mechanisms, but it is at an even earlier stage and has a slightly different focus from MCP, according to Olliffe.
"AGNTCY is not really 'baked' or in wide use," he said. "It's more of a proposal/prototype specification for AI agents to interact with each other. MCP is more focused on enabling an AI tool, app or agent to discover and access commonly used APIs or local commands."
Beth Pariseau, senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism covering DevOps. Have a tip? Email her or reach out @PariseauTT.
