The Protocol That Silicon Valley Called Broken Is the One BFSI Is Building On
Silicon Valley is walking away from MCP for the same reason financial services are adopting it.
In March 2026, Perplexity CTO Denis Yarats stood at the Ask 2026 developer conference and made an announcement. Perplexity, which had shipped its own MCP server just months earlier, was moving away from the Model Context Protocol.
The reason was that MCP tool schemas were consuming up to 72% of the company's available context window space before the agent processed a single word of user input, according to the company.
The authentication model, which requires each MCP server to manage its own credentials, added friction at every connection point. For a consumer AI product competing on speed and responsiveness, the overhead was indefensible.
Within days, Y Combinator president Garry Tan offered a blunter verdict, describing MCP as something that "sucks honestly," and published an alternative CLI-based workflow.
A protocol that had seemed unstoppable with 97 million monthly SDK downloads, adopted by OpenAI, Google, Microsoft, and AWS, donated to the Linux Foundation in December 2025, was suddenly a subject of serious criticism from some of the most credible voices in AI development.
At the same time, in a less visible corner of the same technology landscape, a pattern of adoption had been building in financial services for months. Grasshopper Bank launched what was described as the first MCP server by a US financial institution in August 2025. Bud Financial followed in October. GoCardless launched in February 2026. Nymbus and Meow Technologies both launched in April.
Every one of them chose MCP deliberately, for reasons that have nothing to do with context window efficiency
Anthropic released the Model Context Protocol in November 2024 as an open standard defining how AI agents connect to external tools, data sources, and business systems.
Before MCP, connecting an AI model to any external system required a custom integration like bespoke code, authentication, and maintenance. The M times N integration problem: M models, N tools, and a different integration for each combination.
MCP solves this by providing a single standardized connection layer. Any system that implements an MCP server can be connected to any AI agent with an MCP client.
The promise is interoperability. Build it once and any compatible agent can use it. By early 2026 the standard had achieved broad adoption faster than almost any open protocol in recent memory.
The criticism from Perplexity and Tan is not that MCP is conceptually wrong. It is that MCP's design, which requires the model to receive complete tool definitions, including parameter schemas and response formats, before it can use any tool, creates measurable overhead at the scale of production consumer deployments.
One developer reported that seven MCP servers consumed 98,700 tokens, roughly a third of a 200,000-token context window, before any conversation began.
For companies where context efficiency translates directly to cost and performance, that overhead compounds into a real operational problem.
Why Choose the Same Protocol
The adoption did not happen overnight. It built steadily, starting in the US before the Silicon Valley backlash had even formed. And at every step, the companies adopting MCP were explicit about what they were buying into — not speed, not simplicity, but accountability.
Grasshopper Bank was the first. In August 2025, the digital bank serving startups and small businesses partnered with Narmi to launch what the two companies described as the first MCP server by a US financial institution.
Grasshopper was explicit about why it chose MCP over a direct API approach: the protocol provides a standardised, auditable connection between the AI model and the bank's systems, with controlled access from the start.
The server allows business clients to query financial data through Claude using natural language without custom integrations. The capability was secondary. The governance architecture was the point.
Bud Financial made the same calculation from a different position. The New York-based AI platform for banking data launched its MCP server in October 2025 to give banks, credit unions, and fintechs access to enriched transaction data and customer financial intelligence.
Bud built SOC2 Type 2 compliance directly into the architecture — not as a later addition, but as a design requirement. In banking data environments, compliance is not a feature. It is the product.
GoCardless, a UK-based bank payment company, launched its MCP server in February 2026 to give developers and merchants a natural language interface to its payment platform. The company framed MCP not as a current convenience but as the technical foundation for agentic commerce, the infrastructure layer that will matter when AI agents begin executing payments autonomously, not just surfacing data.
Nymbus made the governance case most explicitly. The US banking platform launched its MCP server in April 2026 with 19 specific banking actions including customer lookup, account management, money movement, and debit card controls.
Every capability came with a corresponding control: token-based authentication, role-based access, PII masking in logs, encrypted connections, and full audit logging. Financial institutions control which tools are enabled, which roles can access them, and where additional approval is required.
Meow Technologies launched the same month, describing its platform as the first agentic banking platform built on MCP, enabling AI agents to open business bank accounts, issue virtual and physical cards, send payments, and manage invoicing.
These are not features these companies added to make MCP safer. They are the reasons these companies chose MCP in the first place.
The Same Properties, Different Conclusions
Perplexity left MCP because the protocol requires structured tool definitions, complete parameter schemas, and a formal authentication flow for every connection. These properties consume context window tokens and create integration friction.
Grasshopper Bank, Bud Financial, GoCardless, Nymbus, and Meow Technologies chose MCP for exactly the same reasons. Structured tool definitions create audit trails.
Complete parameter schemas enforce access boundaries. Formal authentication flows document every tool call an AI agent makes.
The protocol is the same. The use case is not.
In a consumer AI product, overhead is waste. In a regulated banking environment, overhead is evidence. When an AI agent freezes a customer's debit card or initiates a money movement, the financial institution needs to know exactly which agent made the call, under what permissions, at what time, against what data.
MCP's structured tool definitions are what make that documentation possible. A survey found that integration complexity with existing business systems is the top barrier to agentic AI deployment, cited by 67% of enterprise technology leaders.
That is precisely the problem MCP was designed to address. The companies discovering MCP's costs are the ones who needed the benefits least.
What This Means for Fintech AI
The MCP backlash has been loud because the people generating it are visible — startup founders, CTO announcements, Y Combinator leadership.
The financial services adoption has been quieter because Grasshopper Bank, Bud Financial, Nymbus, and Meow Technologies build infrastructure and let the deployments speak for themselves.
But the pattern is evident. As consumer AI companies move toward lighter, faster alternatives like CLI-based tools, single-endpoint APIs, and direct REST calls, the enterprise and regulated sectors are moving in the opposite direction.
The governance infrastructure MCP provides is becoming more valuable, as AI agents begin taking consequential actions in financial systems.
The EU AI Act, enforceable from August 2026, classifies AI systems used in credit scoring, insurance risk assessment, and financial market operations as high-risk, requiring documented human oversight mechanisms.
An MCP server that logs every tool call, enforces role-based access, and masks PII is not a burden for a bank building toward that requirement. It is a head start.