Unknown Title

Key Concepts

• MCP (Model Context Protocol): An open standard that allows AI models to interact with external data, tools, and systems, effectively giving "hands" to chatbots.
• MCP Apps (formerly MCP UI): An extension of the MCP standard that enables MCP servers to render rich UI elements (widgets) directly within the chat interface.
• Context Window: The limited amount of information an LLM can process at once; inefficient UI (like walls of text) wastes this space.
• JSON-RPC: The communication protocol used for bidirectional interaction between the host application, the MCP server, and the UI widgets.
• Sandboxing: A security measure ensuring that UI widgets cannot access the host application's core system without explicit permission.

---

1. The Evolution of MCPs and the "UI Problem"

MCPs have grown significantly in the last year, with over 97 million SDK downloads and 10,000+ active servers. While they allow AI to execute tasks (querying databases, reading files), they historically suffered from a major UX flaw: "Text-in, Text-out."

• The Problem: MCPs often returned data as "walls of text" that users ignored. Forcing an LLM to generate code to render a chart is unreliable, expensive, and consumes valuable context tokens.
• The Insight: As Edos Ruska notes, "If images are worth a thousand words, then good UI/UX is worth 10,000 words." The goal is to shift from "putting chatbots into applications" to "putting applications into chatbots."

2. The Development of the MCP Apps Standard

The transition from a fragmented ecosystem to a unified standard involved a collaborative effort between industry giants and open-source contributors:

• Origins: Started as an open-source fork (MCP UI) by Eido and Lead.
• Industry Alignment: Following concerns about fragmentation when OpenAI introduced proprietary extensions, a collaborative effort (SAP 1865) was launched in November.
• Current Status: As of January 2025, the standard is supported by major clients like Claude, ChatGPT, VS Code, and Goose, with integrations from Figma, Asana, Slack, and Canva.

3. Technical Framework: How MCP Apps Work

MCP Apps function by returning HTML-based UI elements as a special type of "resource" rather than raw text.

• Efficiency: Widgets do not share context with the LLM unless explicitly instructed, saving tokens and reducing costs.
• Communication: The host, server, and widget communicate via bidirectional JSON-RPC.
• UI-Only Tools: Developers can mark specific tools as "UI-only," preventing the LLM from seeing them and keeping the context window clean.
• Styling: Host applications can pass theme information (e.g., dark/light mode) to widgets to ensure they feel native to the environment.

4. Real-World Application: Desktop Commander

Desktop Commander is implementing these features to improve local AI assistance.

• File Previews: Instead of raw text, the app now renders markdown and HTML files in a collapsible, interactive widget.
• Functionality: Users can open folders, copy content with one click, and view rendered formats directly in the chat.
• Future Roadmap: The team is exploring "revertible edits" (allowing users to edit AI output directly in the chat) and "interactive terminal widgets" to allow users to take control of console tasks alongside the AI.

5. Current Limitations and Friction

Despite the progress, the speaker highlights that the implementation is still in its infancy:

• Inconsistent UX: In platforms like Claude, sending a message from a widget often requires the user to manually click "submit" again, which the speaker describes as "useless" and "suboptimal."
• Context Awareness: Currently, the LLM is often unaware of user interactions within a widget (like text selection) unless the user explicitly prompts the AI to "read" the selection.
• Host Control: While the standard provides the framework, host applications still dictate the final user experience, leading to varying levels of integration quality.

6. Synthesis and Conclusion

The introduction of MCP Apps marks a pivotal shift in AI interaction, moving away from passive text-based responses toward active, application-like interfaces. While the standard is currently hampered by early-stage implementation issues—such as poor message handling and limited context awareness—it provides a robust foundation for more seamless human-AI collaboration. The future of AI assistants lies in "seamless" integration, where the AI acts as a side-by-side partner that understands user actions in real-time without needing constant, explicit tool calls.