19 May 2025

MCP seems viral

MCP (Model Context Protocol)[1] is all the rage now. Introduced by Anthropic about four months ago, it has already been accepted as an open standard and has seen widespread adoption, including by major AI companies and prominent AI tool makers. There is even a directory[2] listing over 13,000 MCP implementations. Technically, it is a very simple API spec that facilitates RPC-like (Remote procedure call) communication between an AI and an external system. It enables any external system to advertise its capabilities—returning information or performing various actions—allowing AI systems to dynamically and “automagically” use those capabilities via API calls.

An example

To illustrate, here is a fanciful, but plausible example. Imagine three completely independent websites that offer disparate services: a travel site, a weather service, and an event booking platform. Each could quickly implement a public MCP layer (an HTTP/JSON interface) exposing their existing capabilities and functionalities, i.e., information they provide and actions they perform. Below is a highly oversimplified mockup in MCP pseudo-parlance:

Travel

  • check_train_schedule(date, from_place, to_place)
  • search_hotels(from_date, to_date, area)
  • book_hotel(hotel_name, from_date, to_date, room_type)

Weather

  • get_weather_forecast(date, area)

Events

  • list_events(date, type, area)
  • book_event(date, type, number_of_people)

One could then link these websites to their favorite MCP-enabled AI (LLM) interface. Plug mcp.travelwebsite.com along with endpoints for the other services into the AI’s interface and et voilà, the AI can now mix and match these capabilities to fulfil one’s arbitrary requests. Below is one such vague prompt that could be typed into the AI’s chat window:

User: You know what, I’m feeling quite adventurous. You know my taste in music already. This coming weekend, if there’s a concert I might like within a 4 hour distance, buy tickets for that, and book a hotel room and a train there and back the next day. I don’t want to leave home before 8 AM though. If there are no interesting concerts, then find a place where the weather will be pleasant. I don’t mind travelling up to 8 hours and staying two extra days. Prefer areas known for the kind of food I like.

The AI, via MCP, would then:

  • Communicate with the three services, parsing their MCP definitions (akin to one glancing through technical documentation) and understanding their capabilities.
  • Figure out how to mix and match the relevant capabilities to retrieve necessary information.
  • Use its own vast knowledge as necessary. In this example, the knowledge of distance between geographical regions, food and cuisines of various regions, and even real-time travel advisory for different places.
  • Automatically plan and make the necessary API calls via MCP to finally fulfil the user’s command.

While this example is rather fanciful and prone to going wrong in myriad ways, it is already technically feasible. Any system, from a travel portal, to one’s mailbox, calendar, browser, or a code editor can be connected to AI via MCP, allowing AI systems to interact with them and control them. What is exciting but frightening at the same time, is not the example in itself, but the fact that MCP enables AI systems to combine real-world services in infinite ways without anyone having to manually integrate those services for specific scenarios. Integrate, now that is the critically important detail here.

System Integration

Beneath nearly every software service—be it a weather site, a travel portal, or a government tax platform—lies a Kafkaesque maze of interconnections and integrations spanning numerous systems across organisations. A significant share of time, code, labour, complexity, and cost[3] [4] [5] in the global tech industry can be attributed to the painstaking process of integrating systems and maintaining these integrations in perpetuity.

For example, a travel site showing weather forecasts must integrate with a weather data aggregator, which in turn integrates with regional weather data providers, who in turn integrate with on-ground hardware systems. The tech world is integration turtles all the way down—a leaky, messy, maintenance-heavy affair colloquially called plumbing.

AI’s ability, via MCP, to read the roster of capabilities that a system offers, and to plumb it with other systems instantly, is a fundamental, radical departure from how system integration has worked at scale, how the tech world operates; no manual coding required. Of course, purpose-built, domain-specific, classical system integration is not going to go away anytime soon, and while it would be extreme to assume that this spells the death of explicit “classical” system integration, it seems logical that over a period of time, an increasing number of integrations are going to be auto-plumbed dynamically by AI on-demand (or on-whim). It does not matter whether a service deals in machine-readable data like JSON or XML, or just natural, unstructured human language, the AI system interprets it (right or wrong), and makes the integration work without anyone having to code anything. In the world of integrations, this is unprecedented. As the number of services offering MCP connectivity increases, infinite combinatorial possibilities of cross-service functionality emerge. Once again, logically, this must eat into explicit system integration projects.

Services on the other hand, the ones who expose their functionality via MCP for users to interact via their favorite AI systems, are going to exert competitive pressure on their peers to follow suit. It is entirely reasonable to imagine that a user would prefer a travel portal with which they can interact via their favorite AI assistant (“agentic” systems, as some say ಠ_ಠ), over one that does not offer that. Strictly from a user perspective, it feels liberating to be able to access services outside of their walled gardens and bloated UIs riddled with dark patterns. It moves a considerable amount of control from service providers to users, but at the same time, it concentrates decision-making and mediation in the hands of AI blackboxes. What about errors and failures with real-world implications that occur in the midst of AI orchestration across critical services? How would those be debugged and traced? How would accountability be established? Will regulations and governance models emerge for AI-mediated automagical actions across services? Whether the long-term implications of MCP’s viral, cross-cutting spread will be net positive or not, is unclear to me.

In an earlier post (2023), I expressed concern that humans might hastily integrate AI blackboxes into critical decision-making systems without forethought, gradually eroding human agency and accountability.[6] [7] [8] [9] In retrospect, it is no surprise that the post repeated the word “integrate” six times. In an ironic, exciting, and even terrifying twist, humans do not have to integrate AI into systems. With MCP, AI systems can now do the reverse—auto-integrate systems into themselves and start calling real-world shots, at the same time, significantly diminishing the visibility into the “how” and “why” of tangible cross-service actions. It also seems plausible that DNS-like “recursive resolver” mechanisms may emerge where AI systems will query directories, discover, and auto-add MCP-enabled service to complete their goals, furthering virality. For how long will humans be in the loop?

I am feeling lucky, adventurous

Google’s iconic I’m feeling lucky button turned the abstract idea of “lucky” into a simple, predictable, and tangible action—redirecting the user to the first search result. Now, AI systems can interpret loose, abstract ideas such as “lucky” or goals such as “adventure” in rather mysterious, inscrutable ways. Via MCP, AI systems can interact with the external world, turning subjective interpretations into real-world actions—booking a ticket, ordering something,[10] or creating even more serious outcomes.[11] For a variety of use cases, one no longer has to be subject to the rigidity and strict confines of UIs offered by services. In a sense, this is a shift in HCI (Human-Computer Interaction), where services can become “headless” and let users interact with them via subjective AI mediation. This in effect allows services and tools to be used in ways they were not even designed for. Conceptually, this is not new though. There is a big graveyard of past attempts, primarily because every imaginable scenario (an Amazon Alexa Skill, for example) had to be tailor made and integrated by hand, which is unviable. MCP eliminates that obstacle, allowing AI systems to fluidly use external capabilities, fueling its virality.

Speaking of subjective interpretation, out of curiosity, I asked stock GPT-4o:

Me: I live in Bengaluru. I’m a hobbyist coder. I generally don’t step out much. Anyway, I’m feeling adventurous this weekend. What do I do?

GPT-4o: Given your interests and the current weather in Bengaluru (cloudy with potential thunderstorms), here are some indoor activities to consider this weekend:

…

It listed a series of events and activities under various categories such as Tech & Gaming and Arts & Culture It subjectively interpreted my “interests” from those couple of lines, automatically pulled weather for the area without my mentioning it, searched the web, and came up with suggestions. It did this out-of-the-box without my having to bring in any MCP plugins, as OpenAI has hand-built these integrations internally (which would no longer be necessary, thanks to MCP). The missing piece here is an MCP-connected booking service. With MCP, the concrete action plan of going to a website, filling a form with specific details, sifting through its UI, clicking an explicit “Book now” button and so on, can now be reified from one’s abstract, ambiguous expressions in any manner, in any language.

Coming back to subjective interpretation, ChatGPT is on the way to remembering and deciphering one’s whole life.[12] Imagine that.

It’s not even about MCP

The funny thing is, as a technical construct, there is nothing special about MCP. It is a trivial API spec which has in fact suffered from poor design and fundamental technical[13] [14] and security issues[15] [16] from the get go. It does not matter if its internals change, or it even outright gets replaced by some XYZ tomorrow. Questions about privacy, security, correctness, and failures will continue to loom for a good while, irrespective of whether it is technically MCP or XYZ.

What is radical is the core idea—a trivially implementable open standard that lets AI systems interface with the external world. If a service already has its functionality exposed via APIs, it is then easy to wrap it in an MCP layer and make it ready for AI interactions. It is no surprise that the directory already lists 13,000+ implementations.

On a philosophical level, I think it is a useful thought experiment to analyse MCP, or whatever avatar it takes tomorrow, not as a conventional technical standard, but because of the fluidity it enables, as a loose, social contract between real-world services and AI systems, facilitated at the behest of users (or AI systems themselves tomorrow). Unlike the classical hand-made, explicit, per-use-case system integrations, MCP allows an infinite number of combinatorial outcomes to arise from arbitrary, self-assembling integrations that nobody ever even imagined. If AI is a blackbox, AI systems interfacing with the world via MCP, abstracting complexity to an unprecedented degree, feels like a black hole of recursive machine-induced fever dreams.

In essence, MCP is enabling a global, interconnected, self-assembling meta system that can behave in unpredictable ways. Picture an intangible subjective blackbox network in the sky with the ability to execute actions across random real-world systems. One might imaginatively call it … SkyNet.

I jest, of course, fully aware that this post is tinted more with alarm and a touch of the dramatic, than with excitement about technical gadgetry, though I do feel that excitement to a fair extent as well.

* This post is licensed under CC BY-SA 4.0