"It's API, MCP, and CLI Now" — What Software Needs to Survive the AI Agent Era

Hello, this is Hamamoto from TIMEWELL.

The SaaSpocalypse Shock

In February 2026, SaaS company stock prices dropped sharply across the board. The trigger: the arrival of AI agents like Anthropic's Claude Opus 4.6 and OpenAI's Codex. Fortune magazine named the event "SaaSpocalypse" and reported that over $400 billion in market capitalization disappeared from the market within a matter of weeks [1].

A statement that seems almost prophetic preceded this event. Microsoft CEO Satya Nadella said in a podcast at the end of 2024:

Today's business applications are ultimately just CRUD databases with a bit of business logic [2]

Honestly, when I first heard this, I thought it was an overstatement. But in 2026, those words have taken on real weight. AI agents don't need a beautiful interface. They call APIs directly and execute business logic. The GUI of SaaS products — designed with the assumption that humans would operate them — is having its reason for existence questioned at a fundamental level.

"SaaS is Dead" — the structural change behind that provocative phrase. So how will software change in the AI agent era? And why did I arrive at the conclusion that "it's API, MCP, and CLI now"? Let me dig in.

The End of UI-First

Traditional SaaS was designed with the assumption that humans operate the UI. The core of value lay in UX — how usable and efficient the product could be for getting work done. The flow was simple: humans operate the UI, SaaS processes data, results return to the screen.

2026's reality is different. AI agents become the "users" of SaaS, and humans simply give instructions in natural language. Agents don't look at screens. They don't use mice. What they understand is APIs exchanging data in JSON format. The value demanded from SaaS has shifted from UI usability to API expressiveness and reliability [3].

To frame this as the evolution of enterprise systems: the 1990s were the era of SoR — Systems of Record. ERP and CRM accurately managing customer data and transaction records. The 2010s added SoE — Systems of Engagement — as smartphones proliferated, then big data and AI brought SoI — Systems of Intelligence [2].

And now, in the AI agent era, what's needed is SoK — Systems of Knowledge creation. Not just recording data, but generating knowledge from data and reconstructing business processes themselves. When Nadella said "just CRUD databases," he was criticizing SaaS that remains stuck at the SoR stage.

MCP's Promise — and the Gap on the Ground

To standardize connections between AI agents and external tools, Anthropic proposed MCP — Model Context Protocol — in 2024 [4]. Enabling AI models to call external APIs and data sources in a unified way. Amazon and Microsoft announced support [5] [6], and as of February 2026, over 9,000 plugins have been published — the ecosystem's expansion is remarkable.

But ground-level sentiment tells a slightly different story.

An article on dev.to sharply identified MCP's practical challenges [7]:

Challenge	What's Happening
Configuration fragmentation	Multiple community-built MCP servers proliferate per platform — unclear which to trust. Configuration becomes siloed by tool
API coverage gaps	Many MCP servers are just incomplete wrappers around existing APIs, with frequently missing functions. Autonomous workflows stall mid-process
Token waste	Massive JSON schemas describing tool specifications are sent to the LLM on every API call, consuming context window and driving up costs

On a personal note: I've tried several MCP servers myself and repeatedly hit "wait, this feature isn't here?" moments. The gap between the ideal and the reality is larger than you'd expect when you actually work with it.

In this context, WebMCP — co-developed by Google and Microsoft — previewed in Chrome 146 Canary [8]. It allows websites to directly expose tools to AI agents within the browser, without needing to build a backend MCP server. It's attracting attention as a complement to MCP, but it doesn't fundamentally resolve MCP's own challenges.

The CLI Comeback

While MCP gets all the attention, CLI — the command-line interface, with over 30 years of history — is quietly but unmistakably growing in significance.

The symbol of this is the rise of Claude Code, a terminal-native AI agent. As of February 2026, 4% of all public commits on GitHub are auto-generated by Claude Code — with predictions that this will reach 20% by year's end [9]. A figure that shows software development itself is changing at a fundamental level.

Why is CLI such a natural fit for AI agents? Three reasons [7]:

First, universality. Once configured on a PC, it can be used consistently from IDEs, terminals, and other tools. Like AWS CLI or gcloud SDK, a single binary covers nearly all platform functionality. None of the "this tool needs a separate config file" fragmentation that occurs with MCP servers.

Second, maturity. Thorny problems in computer-to-computer communication — authentication, error handling — have been solved over decades. MFA, SSO — these are long-settled technology in the CLI world.

Third, agent affinity. AI agents don't need to interpret screen context. Structured text information from --help alone is enough to accurately understand how a tool works and execute tasks.

Peter, founder of OpenClaw, flatly states "everything you can do with MCP, you can do better with CLI" [10]. Opinions differ on this, but I find myself substantially in agreement. Node.js creator Ryan Dahl has said "the era of humans writing code is over." Ruby on Rails creator DHH has called writing code by hand a "luxury." The most technically sophisticated engineers are migrating to AI-agent-driven development styles, and CLI is at the center of that [9].

Three Interfaces — They Don't Compete

I've traced the trajectories of API, MCP, and CLI. The real point of "it's API, MCP, and CLI now" is that these three don't compete — they each play a role at a different layer of a complementary architecture.

API is the foundational layer. The contract by which a service exposes its capabilities to the outside world — the official front door for programmatic access. Without a robust and expressive API, nothing on top of it can stand.

MCP is the translator between AI models and diverse tool sets. It converts the infinite variety of APIs with different specifications into a unified format that AI models can understand, standardizing connection. The breadth of the ecosystem determines its value.

CLI is the most natural interaction interface for AI agents. Planning, executing, checking results — the agent's thinking process aligns with command-based sequential operations in a surprisingly natural way.

WebMCP adds another option — a client-side connection mechanism specialized for scenarios where users and AI agents collaborate via browser.

The question being asked is not "which technology is best" — it's binary thinking of the wrong kind. The real question is: how do you make your service agent-friendly? Whether you can provide capabilities in a form that AI agents can easily access and operate will determine whether software lives or dies going forward. On the foundation of API, the connectivity of MCP and the operability of CLI layered on top — this trinity architecture is essential.

What to Actually Do in Practice

The structural change is clear. So what specifically should you do?

First, review your service's API quality. Idempotency guarantees, fine-grained API design, support for asynchronous workflows, detailed audit log provision. These are moving from "nice to have" to "can't survive without" [3].

Second, build out CLI tooling. If your service doesn't have a CLI, start development now. If you already have one, improve help text, add JSON output format support, and simplify authentication flows — all with the assumption that AI agents will be using it.

Third, assess MCP readiness. If MCP is becoming a standard in the ecosystem your service operates in, the response is unavoidable. But rather than building your own server from scratch, also evaluate existing OSS and third-party solutions — and assess cost-effectiveness carefully.

For those selecting SaaS, here's a checklist I'd suggest. The era of judging on UI demos alone is over.

Item	What to Verify
API quality	Is a REST or GraphQL API publicly available? Is idempotency guaranteed?
CLI availability	Is an official CLI tool provided? Is it usable enough?
MCP support	Does an MCP server exist? Is the community active?
Documentation	Is the API reference and CLI tutorial comprehensive?
Auth and permissions	Can service accounts be created for agents? Are API scopes granular?
Observability	Can API usage and audit logs be retrieved via API?

CLI-First Toward the Future

Among the three interfaces, I believe CLI will grow most significantly in importance going forward. CLI is the closest thing to a native language for AI agents. There's no question that API is the foundation, or that MCP plays a role connecting the ecosystem. But for agents executing tasks autonomously, the interface with the least friction and the most expressive power is, at this point in time, CLI — I'll state that clearly.

While writing this article, something struck me: the idea of "CLI-first" is no longer just for engineers. When a product manager is selecting SaaS, when an executive is formulating DX strategy, "does this service have a CLI?" and "is the API sufficiently developed?" are entering the decision criteria. That's the reality of 2026.

"It's API, MCP, and CLI." Evolving your service to be agent-friendly through this trinity perspective — this change is irreversible, and it's a theme that everyone involved in software needs to confront.

Adapting to this structural shift is not something a technical team can complete in isolation. Reviewing API design, building out CLI tooling, determining MCP priority — these are business strategy decisions. At TIMEWELL's WARP program, former DX and data strategy specialists from major enterprises provide hands-on support — from AI agent era DX strategy development through actual API design and system architecture review. "Where do we start to make our service agent-friendly?" Let's work toward that answer together.

---

References

[1] Fortune. (2026, February). SaaSpocalypse: How AI Agents Wiped Out $400 Billion in SaaS Value.

[2] NewsPicks. (2026, February 22). The Real Meaning of "SaaS is Dead" — The Strategy Challengers Should Take. https://newspicks.com/news/16100165/body/

[3] Qiita. (2026, February 17). SaaS Won't Die, but "Design That Assumes Humans Touch the UI" Is Over. https://qiita.com/nogataka/items/e04d1f6f417eec2bab54

[4] Model Context Protocol. (2026, January 26). MCP Apps. https://modelcontextprotocol.io/docs/extensions/apps

[5] AWS Machine Learning Blog. (2026, February 20). Integrate external tools with Amazon Quick Agents using MCP. https://aws.amazon.com/blogs/machine-learning/integrate-external-tools-with-amazon-quick-agents-using-model-context-protocol-mcp/

[6] Microsoft Learn. (2026, February 13). MCP and Foundry Agents. https://learn.microsoft.com/en-us/agent-framework/agents/tools/hosted-mcp-tools

[7] DEV Community. (2026, February 7). The Uncomfortable Truth: Why CLIs Are Still Beating MCP Servers. https://dev.to/mechcloud_academy/the-uncomfortable-truth-why-clis-are-still-beating-mcp-servers-in-the-age-of-ai-agents-4n9f

[8] VentureBeat. (2026, February 12). Google Chrome ships WebMCP in early preview. https://venturebeat.com/infrastructure/google-chrome-ships-webmcp-in-early-preview-turning-every-website-into-a

[9] SemiAnalysis. (2026, February 5). Claude Code is the Inflection Point. https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point

[10] Zhihu. (2026, February 12). MCP Is Dead, 80% of Apps Will Disappear, Vibe Coding Is an Insult. https://zhuanlan.zhihu.com/p/2005607548506621432