The perfect interface is the final piece in the development of any technology that makes it usable to huge masses. Before becoming a chat application, GPT wasn't popular even though the API existed. For MCP, despite the huge surge in popularity in March/April, I don't think we've found the perfect interface yet. Let's examine the existing options: Claude Desktop is really bad to use MCPs with – the app itself is buggy, doesn't handle errors well, and managing MCP Servers is a nightmare. I could disable the Servers and Claude would still not allow me to use the chat if the number of tools is too high (anything over 15-20 tools becomes unusable). It doesn't even come pre-shipped with something basic like a File Server. Cursor and Windsurf initially made MCP popular and useable with their easy server management and smart handling of tool calls, but recently I've found it increasingly hard to use MCP tools with non-Claude models like Gemini 2.5 Pro or GPT 4.1. I keep getting errors requesting me to disable MCP Servers, which defeats the whole purpose.
I've also been building AI agents and using them actively. Agentic systems that don't use MCP don't feel that much different from ones that use MCP. I don't feel the convenience of using MCP. I probably don't have to write the tool but you can just figure out the params and a high level idea of the tool and ask Claude to generate the tool. If you read Anthropic's blog on Context Engineering, you can see they've used Claude to actively build tools and edit descriptions and let Claude figure out the best way to write tools for these AI agents. The convenience is felt in two places though. First is authentication – MCP makes it easier for me to manage authentication. Can you imagine managing environment secrets for every single tool you write? The second place is my terminal, which brings me to why terminal-based interfaces might be the perfect solution for MCP.
Terminal-based software might actually be the answer. It's crazy how after all these years of visual designs and images/videos, we're back to simple text-based interfaces. Every developer hates having to configure their environment and setup before actually coding – I personally hate managing my 5 Python environments. The fewer steps between me and running code, the happier I am. Claude Code is wonderful for using MCP Servers with well-crafted agents excellent at utilizing tools. However, you're stuck with strict rate limits and model companies being greedy with their rules. My biggest frustration is constantly approving requests from Claude Code/Gemini CLI without any ping or sound notification. There's probably a startup opportunity in solving human-in-the-loop communication with agents. Maybe OpenAI has figured it out with their notifications, but we need something better. I really like the way Devin communicates through Slack. That's a nice experience where I can describe the problem and since Slack is where most bug screenshot reports go and tickets are asked to be created, the agent being there works. Similarly, I think ChatGPT Agent is a really good interface for a browser agent.
I'm working on MCP Shell, which I think will be the greatest interface for anyone who wants to experiment with various tools and agents. MCP Shell will allow any LLM to use MCP tools without being restricted by model providers. It won't just be for coding – it'll handle all kinds of agentic use cases. You'll actually use this agent in the background to get work done without constant intervention, and it'll conveniently ping you on any messaging interface when needed. Follow the project at https://github.com/Prat011/mcp-shell or reach out at prathit3.14@gmail.com if you want to contribute.