Your customer doesn’t want your agent, they want your MCP

A little over nine months ago, I wrote Don’t Build Another Chat Box into your Product because I kept seeing people treat a chat box as their AI strategy. My advice today has a similar ring: no one wants to use your agent. I don’t want to open someone else’s chat box and explain myself again when I already have an agent that knows what I’m working on. I want to use my agent to interact with your services.

Agent-to-agent communication isn’t really part of how I work today. Maybe it will be, maybe not, but my agent talking to your API or MCP is already very real. I use Linear, Notion, and GitHub via MCP far more than I use their applications these days. I delegate work to Codex and use those products as context stores, giving my agent somewhere to find the work, the decisions around it, and the code that implements it.

I still like those products and their metaphors. Issues, projects, documents, and pull requests give the work a shape that makes sense to me, and that’s what I pay them for. I don’t think rebuilding those products is a valuable use of my time, either. AI has changed the calculus on buy versus build, but you can still spread yourself too thin. Ask me how I know. I’d rather pay for those capabilities and have my agent work with them than make every piece of my workflow another product I have to maintain.

It’s similar with the App Store analytics service I pay for. I use its MCP to build custom reports through my agent, while the service does the hard work of collecting and packaging the data. I don’t want to rebuild all that grunt work, and I don’t need another agent to tell me what it thinks I should care about. I want the data available to my agent so we can put together the report I actually need.

In some ways, it’s even better if I forget you’re there. If my agent keeps using your service to get useful work done, I’ll happily keep paying. Spending less time in your application doesn’t mean I value the product less; it might mean it’s fitting into my work better. The thing I’m paying for is still doing its job, even when I’m not looking at your interface.

That’s my bet as a builder at Tariff Code Compliance, too. Our website functions more like a showcase for the data and capabilities we expose through MCP. I want someone to look at it and think, “Wow, you have that?” or “You can do that? Let me go use their MCP.” The website gives them a way to see what’s possible, then they can take those capabilities back to the agent they already use and put them to work with the rest of their context.

For the concrete version of that workflow, our Claude and ChatGPT guides walk through connecting and researching a product with your agent. Explorer is the browser path for searching the evidence yourself.