MCP Apps Explained: Interactive UI Inside Claude

Until now, Claude could tell you your campaign CTR was 3.2%. With MCP Apps, it can show you the chart — and let you click, filter, and drill down without ever leaving the conversation. MCP Apps are the biggest shift in how Claude presents information since the platform launched, and they change what "talking to an AI" actually means.

What Are MCP Apps?

MCP Apps are interactive UI components — charts, dashboards, kanban boards, data tables, forms — that MCP servers can render directly inside Claude conversations. They launched in January 2026 as part of Anthropic's rollout of the MCP 2025-11-25 specification.

The key word is interactive. This isn't Claude generating a static image or pasting in a table. It's a live UI element: you can click buttons, apply filters, update records, and approve actions — all from within the chat thread. The conversation becomes a workspace.

If you've used Notion AI or a Slack bot that renders a formatted card, you have a rough mental model — but MCP Apps go further. They're not pre-templated widgets. They're dynamically generated UI from any MCP-compatible server, which means the range of what can be rendered is only limited by what servers choose to build.

How MCP Apps Differ from Regular MCP Tool Responses

To understand MCP Apps, you need to understand what came before them.

When a regular MCP tool runs — say, a tool that fetches your latest Amplitude analytics — it returns data. That data might be JSON, a list of numbers, a structured object. Claude receives it, interprets it, and narrates it back to you: "Your campaign had a CTR of 3.2% last week, up from 2.8% the week before."

That's useful. But it has real limits:

• You can't see the trend line — only Claude's verbal summary of it
• You can't filter by segment without typing a new message
• You can't act on the data — approve a budget, reassign a task — without another round-trip
• Complex datasets lose nuance when translated to prose

MCP Apps eliminate those limits. Instead of returning raw data for Claude to narrate, the server returns a UI component specification. Claude.ai renders that spec as an actual interactive element inside the conversation. You see the chart. You interact with it. Claude can still add commentary — but now it's supplementing what you can see, not replacing it.

Launch Partners and What Each One Does

Anthropic launched MCP Apps in January 2026 alongside a curated set of integration partners. Each demonstrates a different category of interactive UI.

Amplitude — Analytics Dashboards

Amplitude's MCP App renders live analytics dashboards inside Claude. You can ask "Show me my top 5 events by volume this week" and get a real bar chart, not a bullet list. Filters persist — narrow by platform, date range, or user cohort directly in the rendered component. For product teams who live in Amplitude, this means analysis conversations happen where analysis context already is: in Claude.

Asana — Task Boards

Asana's MCP App renders kanban-style project boards inside the chat. You can view task status, reassign work, update due dates, and mark items complete without switching tabs. When you say "What's blocking the Q1 launch?", you get a filtered board view of blockers — not a list of task titles you then have to go look up in Asana.

Figma — Design Previews

Figma's MCP App embeds design previews directly in Claude conversations. Ask Claude to "show me the latest designs for the onboarding flow" and the actual frames appear inline. You can zoom, inspect, and navigate between frames. For design reviews that happen in chat — increasingly common on distributed teams — this removes the constant link-sharing dance.

Other Integration Categories

Beyond the headline partners, the MCP Apps framework supports any server that implements the UI spec. Early community builds include:

• Data tables — spreadsheet-like grids with sort and filter, sourced from databases or CSVs
• Approval forms — structured forms that collect a response and route it back to the server
• Timeline views — Gantt-style project timelines for project management tools
• Map embeds — geographic data visualized in an interactive map component

How MCP Apps Work Under the Hood

The mechanism behind MCP Apps is an extension to the existing MCP tool response format. When an MCP server wants to render a UI component, it returns a response that includes a special content type alongside (or instead of) the usual text or data payload.

In simplified terms, a response might include a block like:

{
  "type": "ui_component",
  "component": "chart",
  "data": {
    "chart_type": "bar",
    "labels": ["Mon", "Tue", "Wed", "Thu", "Fri"],
    "series": [
      {
        "name": "Page Views",
        "values": [4200, 3800, 5100, 4700, 6300]
      }
    ]
  },
  "actions": [
    {
      "label": "Filter by segment",
      "tool": "amplitude/filter_chart"
    }
  ]
}

Claude.ai sees this content type, recognizes it as a renderable component, and passes it to the interface layer for rendering. The actions field is particularly important: it defines what the user can do from within the component — and maps those actions back to specific MCP tools. When you click "Filter by segment", it's invoking amplitude/filter_chart behind the scenes.

This architecture means MCP Apps are stateful and interactive — not one-shot renders. They can call back into the server to fetch updated data, submit forms, or trigger downstream actions.

For a full technical breakdown of the spec changes that enabled this, see our MCP November 2025 spec explainer.

How to Use MCP Apps Today

Getting started with MCP Apps requires a few pieces to be in place. Here's the step-by-step.

Step 1: Use Claude.ai

MCP Apps currently only work in Claude.ai — the web interface at claude.ai. Claude Desktop does not yet render MCP App components. Open Claude.ai in a modern browser (Chrome, Firefox, Edge, Safari all work).

Step 2: Connect a Compatible MCP Server

In Claude.ai, navigate to Settings → Integrations (or the MCP Servers section). You'll see a list of available integrations. Servers that support MCP Apps are typically flagged as "App-enabled" or show a UI icon in the listing. Connect the server you want — Amplitude, Asana, Figma, or another compatible integration.

Step 3: Authenticate

Most MCP servers require you to authenticate with the underlying service. Follow the OAuth flow for your chosen integration. This is a one-time step — your credentials are stored securely and reused in future sessions.

Step 4: Start a Conversation

Just ask. You don't need special syntax. "Show me my Asana tasks for this sprint" or "Pull up the Amplitude dashboard for our iOS app" will invoke the appropriate server and, if it supports MCP Apps, render the component inline. If the server doesn't support MCP Apps yet, you'll get a regular text response instead — no errors, just the old behavior.

Step 5: Interact

Click filters, update records, scroll through timelines — directly in the rendered component. Actions that modify data (reassigning a task, approving a request) will typically ask for confirmation before executing. You're in the conversation the whole time.

Current Limitations

MCP Apps are powerful, but they come with real constraints you should know before building workflows around them.

Claude.ai Only

The biggest limitation: MCP Apps only work in Claude.ai. If you use Claude Desktop, the Claude API directly, or any third-party Claude integration, you won't see rendered components — you'll get standard tool responses. Anthropic has signaled that Desktop support is coming, but hasn't given a timeline.

Server Support Required

MCP Apps are opt-in on the server side. The server has to explicitly implement the UI component spec in its responses. Most existing MCP servers don't yet support this — they return standard data responses. Only servers specifically built or updated for MCP Apps will render interactive components.

Component Vocabulary Is Still Early

The current spec defines a limited set of renderable component types. Charts, tables, forms, and basic boards are supported. More complex UI patterns — multi-step wizards, rich text editors, embedded video — aren't in the initial release. Expect the component library to grow as the ecosystem matures.

No Offline or Export

Components are live views, not exportable artifacts. You can't save an MCP App component as a PDF or embed it outside Claude.ai. If you need a static export, you'll still need to use the native tool (Amplitude's export feature, Asana's reporting, etc.).

What's Coming Next

MCP Apps represent a direction, not just a feature. The trajectory is toward Claude becoming a true ambient interface — a place where work actually happens, not just a place where you get answers about work.

Several developments are already in progress or strongly implied by the spec:

Claude Desktop Support

Desktop rendering is the obvious next step. The component spec is client-agnostic — Claude.ai renders it today, but the same server response could be rendered by Desktop, mobile apps, or embedded Claude instances in third-party products. It's a matter of client-side implementation, not a protocol change.

Richer Component Types

The current component vocabulary will expand. Anthropic and the community are already working on additional component types: rich document editors, multi-panel layouts, embedded media, and real-time collaborative components that multiple users can interact with simultaneously.

Cross-Server Composition

The longer-term vision is components that draw from multiple servers at once — a project dashboard that pulls tasks from Asana, metrics from Amplitude, and designs from Figma into a single unified view. This requires coordination between servers and a more sophisticated rendering layer, but the architecture supports it.

To understand where MCP Apps fit in the broader arc of the protocol, our MCP version history guide traces every major change from the initial release to the present day.