Everyone wants to make things agent-friendly now, so are MCPs better than CLIs?
When Claude Code started hitting limits, I began using Kimi (with OpenCode) alongside Claude. Only then, I realized the portability issue: I needed to set up MCP again on OpenCode, and again for Codex.
While MCPs are tied to the integration layer, CLIs are tied to the operating system. So if you build a CLI, you are betting on something much simpler, that the agent can run a shell command. 😁
A few days ago I asked my friend Marco why Struere, his agent management platform, had a CLI instead of an MCP. He said:
The problem with MCPs is that you add an extra tool to the agent, which by definition makes it work worse. A CLI is different. It uses the tool the agent already had, and you just give it context on how to use that tool.
When an agent already has a bunch of tools, adding one more means that the model now has one more thing to choose from, one more schema to understand, one more tiny decision before it gets to the actual work. You might think you are giving the agent more power, and sometimes you are, but you are also giving it more surface area to reason about. (And where to invest more tokens 😬)
Claude and most of the frontier models have gotten much better at MCP tools, and when the integration is good, it does feel nice because the output is structured. The model does not have to read a messy terminal response and guess what matters. I do not want to flatten the argument into "MCP bad, CLI good," because that is not what I believe.
I still think MCPs are useful when the model needs structured context inside the conversation. I just do not think MCP should be the automatic answer because the product is meant for agents. A good CLI covers a lot of the same use cases with less setup and less context overhead.
If I were building a dev tool today, I would start with the CLI. Then add the MCP.
In my case with Experiwall, that makes A/B testing easier for websites. We still shipped an MCP because some workflows work better when the agent can inspect experiments, variants, and results directly inside the conversation. But I've also added CLI support, so you don’t have to reconnect everything if a developer changes agents (as was my experience 😅).
But if I had to choose which interface gets used more a year from now, I would choose the CLI.
Top comments (1)
And of course CLI feels better when the tool should work across UI based platforms (i.e. the Claude desktop app)