The breakneck development of artificial intelligence tools and applications is visible in the rapid emergence every six months or so of “the next big thing”. Generative AI was the big thing two years ago, agents were the big thing last year and this year it’s the turn of Model Context Protocols (MCPs).
Attracting the “game changer” soubriquet, it was released in 2024 and this year really does appear to be doing what the hype machine says; it’s changing the way many organisations play their game.
Anthropic, the developer behind the Claude large language models (LLMs), created MCPs as an open-source connectivity standard. The company reasoned that existing AI models and agents, though sophisticated, can only work with custom integration of each data source, a time-consuming, expensive and inflexible proposition for organisations such as financial institutions, which frequently onboard data feeds and AI tools and agents.As an open-source technology, MCPs can be developed in-house by organisations without the need for a licence from Anthropic, a factor that has enabled their spread like wildfire across the financial service industry.
“MCP represents a fundamental shift in how enterprises will scale AI, moving from fragmented, custom-built integrations to a single, standardised interface that connects governed data to any compatible agent,” said Dwarak Rajagopal, VP of AI engineering and research at Snowflake, which added a Managed MCP Server to the version of its Cortex AI engine that was created specifically for financial institutions.
“This means our customers can move from AI pilot to production faster, without sacrificing the security and compliance controls their business demands,” Rajagopal told Data Management Insight. “As the ecosystem matures, MCP will become the connective tissue between enterprise data and AI, and the organisations that adopt it early will have a structural advantage in how quickly they innovate.”
New Wave
In the past year, a wave of vendors have built their own MCPs or have provided the means for the technology’s integrations. Alkymi has created an MCP for its private credit investment client base; Fitch Solutions has built one for its Nexus ratings data service; and Arcesium has updated its Aquata AI platform and integrated MCP connectivity.
At its core, an MCP is a layer of instructions for LLMs and agents, said Thomas Li, chief executive of US-based Daloopa, which uses AI agents to build financial models for hedge funds and research houses.
“An MCP is a very fine-tuning set of instructions for the LLM to know how to use your database,” Li told Data Management Insight. “If you didn’t have an MCP and you just exposed your core database to a LLM, it’s the equivalent of if I tomorrow said, ‘Hey, you can come and take a look at the database and our database has 150 million rows of data in it, and you’re like, what am I going to do with this?’.
“If we took the same set of instructions and we codified it into literally an instruction manual, a guidebook, and we handed it over to the large language model, then now we have an MCP.”
Connectivity Better Communicated
MCPs provide contextual richness to AI queries, enabling LLMs to better understand instructions and orchestrate a response among agents. In concept, they are very simple; using text instructions, they allow LLMs to communicate better with databases. Depending on how complex the need, an MCP can provide the means for multiple connections to proprietary, open source and third-party databases.
They broaden what data engineers call the context window of LLMs, meaning that they give models a deeper understanding and, in effect, longer memory of a conversation and interaction with agents, tools and data.
Users and applications, therefore, don’t need to continually remind the model of the context within which their queries lie.
Li illustrates this with a motoring metaphor.
“Imagine if you were driving a car and you had to read the instruction manual every single time before you drove it; by the time you start the car, you’re tired, your brain isn’t operating at 100 per cent because your brain’s context window is getting filled up by the instruction manual,” he explained.
“But if I design a car so intuitive that you don’t need an instruction manual, you don’t have that brain load, so your context window stays open.”
API Customisation Drag
Even AI has capacity limitations and the context window is a hog of that resource. By reducing a model’s need to re-familiarise itself with the context to the task at hand, it frees capacity to do more work, faster and at greater scale.
MCPs have the potential to revolutionise the way financial institutions operate by eliminating the need to build custom API connectors between data and the tools and services that use it. That reduces the time it takes to develop new products and solutions, lowers maintenance costs. It also helps boost system resilience by better managing authentication and access controls.
Peter Kohler, head of product specialists at Fitch Solutions, which has just integrated MCP capabilities into its credit and ratings data provisions, characterises the technology’s effectiveness as being “the USB-C for data”.
“The way it would ordinarily be done today is through the API, which means that with every single source of our data or competitors’ data – anybody’s data – I’ve got to do programmatic work for each API and for each endpoint,” Kohler told Data Management Insight. “MCP does away with that.”
MCPs are highly adaptable. As they provide instructions through text, they can be directed to connect any database – and the applications and tools that use it – to any other database and tools.
Organisations can, then, apply them to a potentially limitless number of use cases.
“The adoption across organisations has been strong, with many claiming that they can now move from pilot to production faster, and connect their data in a way that meets both their compliance and security standards,” Dwarak Rajagopal, VP of AI engineering and research at Snowflake, told Data Management Insight.
“These benefits extend across industries, including highly regulated sectors, such as financial services, where organisations can accelerate AI adoption while maintaining the governance, security, and compliance controls required for sensitive data and critical business processes.”
Adaptability Benefits
The business benefit of what MCPs can do is enormous, said Christopher Sparke, chief commercial officer at Fitch Solutions. The company’s MCP product, Nexus, has been designed to enable clients to quickly access its vast credit ratings and research database within the Fitch environment.
“Integrating [multiple data sets] so that there’s a true reflection you can get from that holistic capability together is incredibly hard; regardless of what most firms say, it’s actually pretty difficult to do,” Sparke told Data Management Insight.
“Now, if I’m the Nexus user, I could say, ‘give me a top-down risk perspective of the autos industry, including a regional industry breakdown, country of risk breakdown, ratings breakdown, and market opportunity breakdown’, and you will get one answer in summary that’s then going to tell you what the insight and data is at the country and the country of risk from a geographic perspective or the geopolitical perspective.
“It brings multiple potential perspectives together.”
Next Big Thing?
As any successful artist or sportsperson will attest, it’s often harder to stay at the top than it is to get there: already, MCPs’ day in the sun may be about to dim. Well-known, and vocal, venture capitalist Garry Tan recently declared MCP as “trash” and several AI developers abandoned use of the protocol with damning parting words.
Li also doesn’t expect the protocols to be pre-eminent for very long, arguing that the need for them in the first place will lead to their inevitable replacement by the next big thing.
“An MCP is a set of instructions on top of an API and the reason you need to give an LLM those instructions is because it doesn’t know how to use your database – it’s just not smart enough,” Li said.
He added that “sophisticated people” are not using MCPs because they see them as a temporary technology.
“But over time, as the models get smarter and smarter, you can make the argument that you don’t need an instruction manual.”
Li envisions a near future in which developers will build models and databases in a way that enables APIs to provide connections between them without the need for an instruction manual.
In the meantime, MCP mania continues apace. For instance, Snowflake recently announced its planned acquisition of MCP-focused governance and observability specialist Natoma to connect its Cortex Agents, Snowflake Intelligence, Cortex Code, and other AI platforms with enterprise systems. And late last year, AWS gave the protocol its continued vote of confidence after the Linux Foundation took the technology under its wing.
Fitch, too, has plans to make more of its data accessible via MCPs.
“It allows us to deliver incremental and supplementary use case value to customers that may have struggled to get that value before this technology was in place,” said Sparke.
Subscribe to our newsletter
