I've built both for the same tool. Anything where you need to see the exact numbers and not the model's summary of them and anything that has to be limited no matter what the agent asks for stayed in UI. And everything exploratory went with mcp.
I could see there being utility along side this that's sort of a checksum for what you see in the agent so you have verification that what you're seeing is exactly what the upsteam sent.
I built an MCP for persistent agents in Claude Code. It has to store run memories in SQLLite. I built a pane in Claude Code for the user to monitor and control agent runs - approve actions etc. The user can also interact with memories. That's the main UI. So it's Claude doing the work but it still needs a UI for user convenience. Everything could be done via claude and prompting, but for actions that are so routine and mechanical, a UI just makes sense.
both are needed, for different reasons.
I built GigaMail, a local email client with an MCP server. Agents can read, search and draft emails and other stuff through MCP, but sending, deleting, moving emails ot searching files requires approval outside the agent.
That separation taught me something: the UI isn't just for onboarding or less technical users. Sometimes it's the security boundary, much easier to confine the agent inside it. Either cases the agent shouldn't be able to authorize its own actions.
https://github.com/adecubed/gigamail
UI is must-have if your users want any sort of precision, determinism, or certainty. UI is faster than giving complete-enough instructions to an LLM for tasks below a certain complexity level — and that threshold rises with your user’s skill and familiarity with the task and how opinionated they are about how it should be done.
Thanks. I tend to agree but interesting that another person commented that UI is for "initial interactions" and MCP is for "power users", which is the opposite of what you are saying I think.
I'm experimenting with this balance now -- I built and launched a small analytics service (basic web tracking) that uses the web UI for sign-up through Cloudflare Access and provides minimal instructions on connecting your agent to MCP.
There's minimal reporting post-login about the sites you're tracking, but site setup, configuration, and reporting is all done through MCP with your agent. It's working for my minimal needs right now (traffic over time, plotting visits on a map) and is a good way to figure out how I want to do reporting rather than building a more general purpose reporting interface.
First of all, it comes down to product positioning. Most of the apps I build are for my own work or collaborative reasoning tasks. Sometimes I build not just apps, but also web services and assorted small tools. My philosophy is to minimize UI as much as possible and keep interfaces lean. Anything that can run in the background stays out of the user’s view — that’s the product perspective. But when I build an app to accelerate my own research work, I tend to go for more complex UI layouts, laying out the full research workflow visibly on the interface.
Developers should provide a documented CLI for all their programs. Anything that is clickable should be MCP-able.
I do not think applications will be only headless. MCP will be an alternative user(agent) interface not a full replacement.
Do you really think that we have to get a LLM subscription just to use a program, that probably has a subscription.
More over, 5 years ago , software developer needed : wikipedia, stackoverflow an and IDE to develope software, now suddenly they cannot live without at $200 subscription to develop a variation of ( video games, custom harness, custom RAG, custom CRUD).
Someone said in 2 years, most developers skill will shrivel so much that the next computer will come with only (yes / no button)
I've built both for the same tool. Anything where you need to see the exact numbers and not the model's summary of them and anything that has to be limited no matter what the agent asks for stayed in UI. And everything exploratory went with mcp.
Would agree here. When you want to see "the real thing" and know it's unmodified.
Similar to this, I built a tool that lets you see what something will look like "once live" in a way that the agent cannot modify it once you've approved it: https://github.com/scottfalconer/verified-human-approval
I could see there being utility along side this that's sort of a checksum for what you see in the agent so you have verification that what you're seeing is exactly what the upsteam sent.
To me MCP is just another User Interface, though the "user" is most likely an agent. A user here being something working towards it's own goal.
The key is that whatever logic / operations you have should be decoupled from any notion of interface / channel / modality in general.
I built an MCP for persistent agents in Claude Code. It has to store run memories in SQLLite. I built a pane in Claude Code for the user to monitor and control agent runs - approve actions etc. The user can also interact with memories. That's the main UI. So it's Claude doing the work but it still needs a UI for user convenience. Everything could be done via claude and prompting, but for actions that are so routine and mechanical, a UI just makes sense.
both are needed, for different reasons. I built GigaMail, a local email client with an MCP server. Agents can read, search and draft emails and other stuff through MCP, but sending, deleting, moving emails ot searching files requires approval outside the agent. That separation taught me something: the UI isn't just for onboarding or less technical users. Sometimes it's the security boundary, much easier to confine the agent inside it. Either cases the agent shouldn't be able to authorize its own actions. https://github.com/adecubed/gigamail
UI is must-have if your users want any sort of precision, determinism, or certainty. UI is faster than giving complete-enough instructions to an LLM for tasks below a certain complexity level — and that threshold rises with your user’s skill and familiarity with the task and how opinionated they are about how it should be done.
Thanks. I tend to agree but interesting that another person commented that UI is for "initial interactions" and MCP is for "power users", which is the opposite of what you are saying I think.
I'm experimenting with this balance now -- I built and launched a small analytics service (basic web tracking) that uses the web UI for sign-up through Cloudflare Access and provides minimal instructions on connecting your agent to MCP.
There's minimal reporting post-login about the sites you're tracking, but site setup, configuration, and reporting is all done through MCP with your agent. It's working for my minimal needs right now (traffic over time, plotting visits on a map) and is a good way to figure out how I want to do reporting rather than building a more general purpose reporting interface.
First of all, it comes down to product positioning. Most of the apps I build are for my own work or collaborative reasoning tasks. Sometimes I build not just apps, but also web services and assorted small tools. My philosophy is to minimize UI as much as possible and keep interfaces lean. Anything that can run in the background stays out of the user’s view — that’s the product perspective. But when I build an app to accelerate my own research work, I tend to go for more complex UI layouts, laying out the full research workflow visibly on the interface.
MCP - power users
UI - initial interactions with the product
Developers should provide a documented CLI for all their programs. Anything that is clickable should be MCP-able.
I do not think applications will be only headless. MCP will be an alternative user(agent) interface not a full replacement.
Do you really think that we have to get a LLM subscription just to use a program, that probably has a subscription.
More over, 5 years ago , software developer needed : wikipedia, stackoverflow an and IDE to develope software, now suddenly they cannot live without at $200 subscription to develop a variation of ( video games, custom harness, custom RAG, custom CRUD).
Someone said in 2 years, most developers skill will shrivel so much that the next computer will come with only (yes / no button)
UI when there’s paying customers.