%20(1).svg)
My mom thought AOL was the internet. Then she thought Internet Explorer was the internet. Then Google. Each time, I'd have to explain that the thing she was using wasn't the internet itself—it was just the door.
Now, the door for developers is the CLI, and the door for consumers is Claude and ChatGPT.
You ask a question or describe an outcome, and the agent produces it. In the past, every software application was accessed through a UI or API. But how do agents interface with software?
MCP—Model Context Protocol—is how agents interface with software. Like HTTP for browsers or SMTP for email, it's a protocol that defines how a calling agent connects to a system, reads its state, and writes back to it. Anthropic introduced it in late 2024. Competing protocols have emerged, but they solve different problems. Google’s A2A handles agent-to-agent coordination, IBM’s ACP handles local orchestration. MCP is the standard for tool and system access, and it's won that category fast: OpenAI, Google, GitHub Copilot, and Cursor have all adopted it, and the public server registry has grown nearly 8x in the past year.
Every major platform shift forces software companies to rebuild their surface area. With the internet, you needed a cloud strategy. With the iPhone, you needed a mobile strategy, and if you didn't have one, you watched your users disappear into apps that did. Those shifts also created entirely new categories: SaaS, the app store economy, on-demand everything. The agentic shift is the same forcing function. If your software can't be reached by an agent, it won't be reached.
Salesforce just made the most dramatic bet in its 27-year history. At TDX in San Francisco, they unveiled "Headless 360," a sweeping initiative to expose every capability in the platform as an API, MCP tool, or CLI command. The next generation of Salesforce usage won't happen in a browser. It'll happen in ChatGPT, Claude, OpenClaw, and whatever comes next.
At Ramp, Teddy Riker wrote recently that "the 80/20 has flipped: the new 80% of interaction with software will be through agents." Weekly active users on their MCP grew 10x in three months as more customers started reaching into the product through Claude, ChatGPT, and other agents. His prediction: unless a product can be reached by an agent via MCP, API, or CLI, it won't survive.
Garry Tan, CEO of YC, is a good bellwether for technical trends. Late last year, he posted that "MCP sucks honestly"—too much context window, bad auth. A few weeks later: "I am coming around to the fact that MCP, done right, can be magic." By April 2026, he'd open-sourced GBrain, his personal AI infrastructure, exposing 30 MCP tools with full MCP support.
This shift is moving fast, and the tooling hasn't caught up. It's arguably a bigger disruption than the shift to cloud or mobile. The entire SDLC needs to be rebuilt for the agentic world: tool schema design, prompt capture, evals, quality gates, production monitoring, regression testing, and observability into what agents are actually doing.
The explosion of APIs in the mid-2000s is a helpful analog. Before Postman, testing an API was manual, inconsistent, and painful. Postman made it visual, reproducible, and shareable, and unlocked an entire generation of API-first development. But Postman was just the start. What followed was an entire ecosystem: Swagger for documentation, OAuth for auth, Kong and Apigee for throttling and metering, Datadog for monitoring, and Stripe for API-based billing. MCP needs all of that built from scratch.
MCPJam started as a fork of Anthropic's own MCP Inspector and quickly became the most popular MCP testing tool on GitHub. We used it to build our own MCP integrations at OCV, and it spread organically through our portfolio. Almost every company we work with has ended up using it independently. I haven't seen that kind of unprompted coverage across a portfolio since Stripe and PostHog at YC.
Right now, most MCP integrations ship blind. Teams can't see what's working, what's breaking, or what callers are actually trying to do. Your actual users are invisible to you inside an agent conversation you'll never see. MCPJam breaks that invisibility and gives you visibility into what's happening between the agent and the MCP server.
The front door to the internet has changed before, and it's changing again. This time, the door isn't a browser, a search bar, or an app store. It's an agent.
MCP has won the protocol layer. What gets built on top of it, the tooling, the infrastructure, the developer primitives, is still wide open. Every category that got rebuilt for the API era will get rebuilt again for the agentic era. The teams that don't make that transition won't survive it.
That's why I'm bullish on MCPJam. It has the right wedge, the most popular MCP testing tool on GitHub, already spreading organically through the best development teams, and a long-term vision for what the agentic SDLC needs to become. The front door is changing. MCPJam is helping teams build for what's on the other side.