MCP changed how I think about building tools

Before MCP, every tool I built was an island. A reporting dashboard that you visit. A CLI script that you run. A chatbot that you query. The tool and the AI assistant were separate things — you'd ask the assistant a question, then go open the tool to find the answer, then come back and tell the assistant what you found.

Model Context Protocol inverts that. The tool doesn't sit in a browser tab. It sits in the agent's context, as a set of callable functions. The agent can reach for the tool mid-thought.

I built a Jira MCP server in February and it was the first time I felt the shift viscerally. I was asking Claude about a sprint, and instead of me going to Jira and copying back a list of tickets, the agent just... got them. The conversation didn't stop. The context didn't reset.

What this unlocks

The obvious thing is convenience. The less obvious thing is agent capability compound interest.

Once I had a Jira MCP server, I started thinking about what else could be a tool instead of a manual step. Trello board state. Kwikly's production database schema. My meeting transcripts. My calendar. Each one I wired up reduced the number of times I had to stop a conversation to go fetch context.

The Daily Notes project is the full realization of this: Slack, Jira, Granola transcripts, and calendar events all flowing into a single context layer that the agent can query from any conversation. Six months of signal, always within arm's reach.

The mental model shift

The old model: I use tools, I tell agents what the tools told me.

The new model: agents use tools, I describe what I'm trying to accomplish.

That's a small sentence. It's a big change in how I scope and build things. Everything I build now, I ask myself: should this be a page you visit, or a server that an agent can call?