Almost everything my AI agent hands me is text. Paragraphs, a bullet list, sometimes a table. For a quick question that is fine. For a meal plan, a contract deadline or a sales pipeline it takes a lot of space and a lot of my time, because I end up reading a description of something I could have looked at directly.
This article is also available in German: MCP Apps: Wenn der KI-Agent eine Oberfläche zeigt.
Mike Cannon-Brookes Named the Problem
I heard Mike Cannon-Brookes, one of the founders of Atlassian, say recently that 40 years ago we worked in a terminal, and now we are back in the terminal with our agents. He thinks that cannot remain the solution.
I know what he means. A contract deadline in a terminal is a line of text. I want to see it next to the other deadlines.
What an MCP App Is
MCP stands for Model Context Protocol, an open standard that lets an AI agent call outside tools and systems. My agent is SeibertGPT in the Weldall Client, our desktop app, and both demos below run there.
An MCP server used to give my agent text. With MCP Apps it can hand over an HTML interface as well, and the agent renders that interface inside the chat. I open it as if it were a website, and it is one: HTML elements presented in my agent window. I click around, and I can keep typing to the agent at the same time. The agent still knows which meal or which contract I am on. The text answers do not disappear. The app sits next to them.
Video 1: What’s for Lunch?
I ask my agent what’s for lunch. The app behind it was vibe-coded by my colleague Michael, which means he built it by describing it to an AI rather than writing the code himself, and he runs it. The agent calls it, and the meal plan appears in the chat.
I get the meals for the next couple of days, with pictures and prices, and I can see where I am registered. Our kitchen was closed that day, so we got pizza. If the meal came from our own kitchen, I could ask to have it packed or kept warm because I am coming later. I could also pass it on to a colleague.
Most of this was possible through an AI agent before. What is new for me is that a website came into my agent, and I can work with it while I keep talking to the agent.
Video 2: The Weldall Client
The second video is the one I would show a customer. I download the Weldall Client from gpt.seibert.ai, install it, log in and confirm. Under Connections I can connect Jira, Confluence, Gmail, Outlook, HubSpot and SharePoint. For the demo I connect the demo workspace, which holds sample data for a made-up company. I ask my questions in English, but the demo UI is in German, so the screenshots below are too.
Three things happen in the video.
A contract radar that shows the deadlines. I open a contract and ask what to do next. The agent pulls more information about that contract and proposes the next steps.
The sales pipeline. It arrives as a funnel. I can filter it by employee, compare one person’s pipeline with the whole team’s, and drill into single opportunities. Then I ask the agent to schedule a meeting with the customer, Sägewerk Fichtelhain GmbH, a fictional sawmill from the demo data. It works from the same context.
The ticket list. Ten open tickets. One of them is at risk of missing its SLA, the response time we promised the customer, and the app highlights it before I go looking for it.
Everything in the demo is sample data, including the pipeline numbers.
Loading content...
Who Gets to See What
The systems and the skills behind these apps, meaning the tools and automations the agent calls, run with each user’s own permissions. IT decides what I can see, and the agent stays inside those limits. SeibertGPT is GDPR-compliant. Many of our customers are European companies, so that matters to us.
Try It Yourself
The lunch app is internal, but you can try the Weldall Client yourself at gpt.seibert.ai. More on Weldall: seibert.link/weldall.
If you want the bigger picture, I wrote about the state of AI at Seibert Group and about why I would rather hire agents than buy more software.
I have enjoyed ordering lunch through my agent this week. That is why I recorded the videos.