Leapfrog your own roadmap: what APIs teach about MCP

The trajectory problem

Every growth plan I have seen in a large, regulated business starts with the same hidden assumption: we keep going the way we are going.

I think that is realistic. Size, scale, strategy, regulation and shareholders all set a business's direction of travel and its velocity. The bigger and more regulated it becomes, the more it is shackled to its current trajectory. A new product needs a business case, a risk review and a place in a roadmap that was fixed a year ago. In collections, lending and financial services, where I have spent much of my career, that is how it should work. It also means the business cannot easily turn.

So how does a business leapfrog its own trajectory, reaching new customers, products and capabilities, without rebuilding itself? The answer has been the same twice in thirty years: connect to other people's speed. The first time the connection was called an API. An API, short for application programming interface, is the doorway one piece of software offers to another. The second time it is called MCP, and I will explain it shortly.

How APIs changed growth for those who used them well

On 16 July 2002 Amazon launched Amazon.com Web Services for developers, free of charge. Jeff Bezos said Amazon was "putting out a welcome mat for developers", and that it could not wait to see how they would surprise it. That much is documented. You may also have heard of a Bezos memo that forced every team to offer its data through service interfaces. The source is a 2011 post by Steve Yegge, who recalled it as "around 2002 I think, plus or minus a year". I could not find the original memo, so I treat it as a well-told story and not a fact.

Salesforce gives a cleaner case of an ecosystem feeding growth. In December 2015 it agreed to buy SteelBrick, a quoting and billing product that its own filing described as "delivered 100 percent natively on the Salesforce platform". A platform is software that other companies build on, as shops build on a high street. The value was about $360 million, and Salesforce Ventures already held a stake. A partner grew on the platform, the owner watched, then bought it. I have not checked how the deal performed afterwards.

Other large companies bought the plumbing, meaning the connecting layer that lets systems talk to each other. These are announced figures:

Each buyer paid for capability that customers were already using. I make no claim about which price was right.

• Google agreed to buy Apigee, which manages APIs, in September 2016 for about $625 million in cash.
• Salesforce agreed to buy MuleSoft, which connects systems together, in March 2018 for an enterprise value of about $6.5 billion.
• Microsoft agreed to buy GitHub in June 2018 for $7.5 billion in stock and promised to keep it an open platform. Microsoft said the deal would add to operating income in fiscal 2020 but dilute earnings per share by under 1% in fiscal 2019 and 2020.
• Twilio agreed to buy SendGrid, an email service, in October 2018 in an all-stock deal valued at about $2 billion.

The platform owner's dilemma

The owner of a platform faces a choice: keep the doors open and let others grow on it, or close them to protect its own growth. Twitter shows the tension. Outside developers built much of what made it useful. In April 2010 it bought Tweetie, a popular iPhone app, and relaunched it as the official app. In May 2011 it bought TweetDeck, reported at about $40 million, a figure Twitter never confirmed.

Then the doors narrowed. In September 2012 Twitter released version 1.1 of its API. Apps that copied its core experience needed its permission once they passed 100,000 user tokens, the digital keys that let an app act for one account. In February 2023 it announced that free access would end on 9 February. Whether that was wise is a matter of opinion. Anyone who had built a business on that access had to start again.

Slack shows a milder version. On 29 May 2025 it tightened the limits for apps sold outside its marketplace, and its updated terms barred bulk export of data and use of that data in large language models, the engines behind AI tools. Salesforce said the changes enhanced privacy and security. Others read them as protecting its own AI plans. I cannot tell you which is right. I can tell you a platform owner can change the terms, and your growth plan should survive the day it does.

What MCP adds

Here is the business consequence first: AI assistants can now reach the software and data you already have, through one common connection, and so can the nimble companies building on top of them.

MCP stands for Model Context Protocol. A protocol is an agreed set of rules for how two systems talk. MCP is an open standard, meaning a published specification that no single company owns. Anthropic published it on 25 November 2024. Think of a common plug, like USB-C, that lets AI assistants connect to your software instead of needing a custom wiring job each time. The project itself uses that analogy. A server, in this sense, is simply a system's side of the plug: the place an assistant connects to.

The pace of adoption is the reason to take this seriously. On 26 March 2025, OpenAI's Sam Altman wrote "People love MCP and we are excited to add support across our products." On 9 April 2025, Google DeepMind's Demis Hassabis said Google would support it in Gemini. On 19 May 2025, Microsoft announced support in Windows 11 and said it was working with the MCP steering committee on security.

On 9 December 2025, Anthropic donated MCP to the Agentic AI Foundation, hosted by the Linux Foundation, a non-profit that looks after shared software such as the Linux kernel. Anthropic, Block and OpenAI co-founded it, with Google, Microsoft, AWS, Cloudflare and Bloomberg as supporters. Anthropic reported more than 10,000 active public MCP servers and over 97 million monthly downloads of its software kits. Those are Anthropic's own figures. The project's governance page describes maintainers who join as individuals, not as companies, which makes it harder for any one vendor to turn the standard into a toll gate.

Two more terms matter to your growth plan. An agent is an AI assistant that can carry out a task for you, not just answer a question. Agentic describes work done that way. Headless software has had its front screen taken away so that other software and AI assistants can use its functions directly: a shop with no shopfront, only a delivery hatch. For growth, that means your capabilities can be reached from anywhere, including by partners and AI.

Why Salesforce chose to open up

I spoke with a highly regarded solutions architect in Silicon Valley who works with Salesforce's large private equity clients. They and their colleagues were surprised by the decision to open up. A platform company with a famously valuable data moat choosing open standards was not what many practitioners expected. That is one practitioner's reaction, not evidence, and it says nothing about how Salesforce sees it.

The public facts are these. In June 2025 Salesforce said its Agentforce platform would include a native MCP client and listed MCP servers from more than 30 partners. In April 2026 it launched Headless 360, a version of its platform with the front screen removed, so that partners and AI assistants can use its functions directly. In August 2026 it announced a partnership with Anthropic, with Salesforce running inside Claude through an MCP server. That product was at pilot stage at launch.

Marc Benioff's reasoning is public. At that launch he said "Probabilistic intelligence alone doesn't run a company, and deterministic systems don't reason." I read that as a plain statement of need: AI models need the governed business data that a system of record holds, and the system of record needs the models. At Dreamforce in September 2026 he added that "AI is creating an interface revolution", meaning staff will reach Salesforce from inside the AI tools they already use.

Note what Salesforce did not do. It tightened Slack's bulk data terms in 2025, and then in February 2026 Slack released its own MCP server and a search API. It opened the door, but on its own terms, with permissions and consent. I call that a sensible way to be open.

The growth playbook: stay close, watch, listen, analyse, partner, then buy

This is the how. You do not need to build every new capability or buy every promising company. My approach has six steps.

Stay close to the nimble businesses and entrepreneurs building on connections to your software. Watch what they build and which customers adopt it. Listen to those customers, because they will tell you what is worth having. Analyse usage and economics. Partner while the evidence is incomplete. Buy only when the evidence is in.

An accretive acquisition is one that adds to earnings per share, or to value, quickly. Microsoft's GitHub deal shows why you need to be precise: it said it would add to operating income but slightly dilute earnings per share at first. Watching first does not guarantee an accretive deal. It does mean you buy with evidence of demand, product quality and the team, not a pitch deck.

What the frameworks say about growth and velocity

Six well-known frameworks, in my reading, support the same playbook.

The BCG growth-share matrix, published by Bruce Henderson in 1970, sorts activities by market growth and share. Its "question marks" have high growth but low share, and BCG's advice is to test them quickly and cheaply and back the few that look like stars. An ecosystem of builders is a field of question marks that others fund.

Open innovation, from Henry Chesbrough's 2003 book, holds that firms should use outside ideas and outside routes to market as well as their own. Build, Borrow or Buy, by Laurence Capron and Will Mitchell (2012), sets the sequence: build if your own assets fit, borrow through a contract or alliance if a partner can supply the capability, and buy if you need broad and deep ties. A connection is a low-cost way to borrow.

McKinsey's Three Horizons, from Baghai, Coley and White (1999), asks you to run today's core, grow emerging businesses and seed options for the future. Outside builders can hold your third horizon until one proves it matters.

Wardley Mapping, created by Simon Wardley, places each component of a service on a scale from new and uncertain to commodity. A partner whose product is becoming well understood, but is not yet widely copied, is a candidate to buy. One already at commodity stage is better rented.

The Rockefeller Habits, by Verne Harnish (2002), are a management rhythm of priorities, data and regular meetings. I mean Harnish's framework here, and John D. Rockefeller only as contrast: Standard Oil owned much of its chain, from barrels to pipelines, while an ecosystem approach does the reverse. The rhythm is what keeps pace. Run the watching as a habit with one owner, a monthly review and a short list.

Two warnings. BCG's 2022 work on ecosystems says fewer than 15% last over the long run, and advises against defaulting to the orchestrator role. Christensen's disruption theory (1997) says small entrants serving overlooked customers can overtake incumbents, which is one more reason to keep them close.

What to do on Monday

Five things, none of which needs a purchase.

• Write down your business's current trajectory: the three things that most limit how fast you can move.
• Name one owner for the watching job, with a monthly slot and a short list of five builders to follow.
• Ask which of your customers already connect AI assistants to your systems or data, and what for.
• Decide in writing what you would never open, and on what terms you would open the rest.
• Set a rule for the buy decision: what evidence of demand, usage and team would make you act.

Where GenWA fits: speed and control together

Moving fast only works if you can see what is connected. GenWA helps businesses see which MCP connections exist, then put rules, approvals and a record around them, through Guard, Evidence and Log. Control applies to AI used through GenWA, and it can run in your AWS or Azure account, or be hosted by GenWA on AWS as agreed. The place to start is the Analyse health check, or an advisory conversation.