MCP Apps gives the Model Context Protocol a new role: it can now support interactive software surfaces inside an AI conversation, not just text responses. The extension is now available on Github, according to the MCP Core Maintainers.
The change matters because many AI tool results are too large, visual or stateful to work well as plain text. Instead of forcing every follow-up action through another prompt, MCP Apps lets tools return interfaces that users can inspect and manipulate directly.
What MCP Apps adds to MCP
MCP Apps is described as the Model Context Protocol's first official extension. Its core idea is simple: an AI tool can return a user interface, and that interface can render inside the chat where the user is already working.
Those interfaces can include dashboards, forms, visualizations and multi-step workflows. That gives developers a way to present tool output in a format that matches the task, rather than squeezing every result into a text summary.
The problem is clearest when a tool returns database queries with hundreds of rows. A model can summarize the result, but a user may still want to sort through the material, compare entries or inspect specific records. With a text-only exchange, each of those actions requires the user to ask again.
MCP Apps changes that pattern. The interface can handle direct interaction, while the model remains connected to what the user is doing. The source describes support for live updates, native media viewers, persistent states and graphical manipulation.
Why interactive AI responses matter
Text remains useful for explanation and synthesis, but it is not always the best container for work. A dashboard can communicate changing metrics more clearly than a paragraph. A form can gather structured input more reliably than a series of typed replies. A visual document viewer can help users move through source material without asking the assistant to restate each section.
The documentation lists several use cases that show where the extension could fit. These include configuration wizards, document review with highlighted PDF sections and real-time server metrics monitoring.
The ext-apps repository also includes an SDK and working examples. The examples cover 3D visualization, interactive maps, document display and music notation. Taken together, those examples point to a broader ambition: AI conversations that can host focused tools without becoming separate applications.
For users, the practical benefit is less friction. A person can ask a question, receive a structured interface and continue exploring the result in place. The assistant does not disappear, but the interaction is no longer limited to written back-and-forth.
Developers get a cross-client path
The MCP Core Maintainers claim that developers can now ship an interactive MCP experience across multiple clients without writing client-specific code. That is a significant promise for anyone building AI tools, because client fragmentation can make interactive features harder to maintain.
The extension builds on MCP-UI and the OpenAI-Apps SDK. An Anthropic spokesperson told THE DECODER that MCP Apps is the open standard built on MCP primitives.
That positioning is important because MCP is already about connecting models to external tools and context. MCP Apps extends that idea from data exchange into interface delivery. The tool is not only returning information for the model to discuss; it can also provide a working surface for the user.
The list of supported and interested platforms suggests the extension is not being treated as a narrow experiment. Claude already supports MCP Apps on web and desktop. Goose and Visual Studio Code Insiders also support it. ChatGPT is expected to follow this week, while JetBrains, AWS and Google Deepmind have signaled interest.
How safety is handled
Interactive interfaces inside chat also introduce new trust questions. MCP Apps addresses that with several containment and visibility measures described in the source.
- UI content runs inside sandboxed iframes with restricted permissions.
- Hosts can inspect HTML before it is rendered.
- Communication uses loggable JSON-RPC messages.
- Explicit user consent may be required for UI-initiated tool calls.
These controls are meant to limit what embedded interfaces can do and make their behavior easier to review. They do not remove the need for judgment. The announcement notes that users should still vet servers carefully before connecting.
That warning fits the broader direction of MCP Apps. The extension makes AI responses more capable by bringing software-like controls into the conversation, but it also makes the choice of connected servers more important. A richer interface can be more useful, yet it also deserves the same scrutiny users would apply to any tool that can act on their behalf.
The bigger shift
MCP Apps points toward a more practical form of AI interaction. Instead of asking a model to turn every tool result into prose, developers can give users the interface that fits the job.
For database results, that could mean a browsable view. For documents, it could mean highlighted sections in context. For monitoring, it could mean live server metrics. For creative or technical data, it could mean maps, 3D views or music notation rendered where the conversation is already happening.
The model still has a role. It can interpret, guide and respond as the user works through the interface. But MCP Apps gives the UI responsibility for tasks that text alone handles poorly.
That is the main significance of the extension: it turns the chat window from a place where tools report back into a place where tool-driven work can happen directly.