MCP Apps Define the Agentic Web's Future

Liad Yosef, who is the co-creator and maintainer of the MCP Apps specification, highlights its role in enabling a unified communication protocol between AI agents and external services.

The standard allows these AI-powered experiences to operate consistently by providing a common language for agents to understand and interact with different web applications, ensuring seamless integration across the AI ecosystem.

The MCP Apps standard offers mutual benefits for all stakeholders, as it allows service providers to send their own UI components directly into chat interfaces. This capability ensures that applications can maintain their distinct brand identity and user experience, even when accessed through an AI agent.

Users, in turn, gain trust by interacting with familiar interfaces, such as Booking.com, within a chat environment. Concurrently, AI agents benefit by acquiring enhanced capabilities without the need to develop custom integrations for every specific action or service they might encounter, streamlining their operational efficiency.

The 'Last Mile' concept addresses the critical interface where increasingly autonomous AI models still require human intervention for final decisions. As AI models improve and perform tasks more independently, humans remain essential at the end of the decision-making chain.

This phase specifically involves UI-driven interactions, such as selecting a seat on a flight or confirming a hotel booking. MCP Apps is crucial in this context, as it ensures that complex legacy services remain accessible and actionable for modern AI agents, effectively bridging the gap between AI autonomy and human-specific choices.

The 'agentic web' signifies a profound shift from the traditional web, where users manually navigate multiple tabs and learn distinct user interfaces for each service. This new paradigm is not merely a web populated by agents but a fundamental change in how websites and browsers function.

Agents can deconstruct websites into 'atoms,' or modular functional components, allowing users to interact with services like Jira or Airbnb without needing to engage with their full, often complex, dashboards. User intent is communicated directly to the services via an agent, bypassing the need for manual UI navigation and streamlining interactions.

This shift liberates users from the inefficiencies of the current web, enabling a more direct and intuitive engagement with online services through intelligent agents.

The emerging vision for the agentic web is centered around a single personal agent that manages tasks across multiple platforms, contrasting sharply with the current fragmented browsing experience. This personal agent will leverage individual context, such as calendar events and user preferences, to proactively assist with various online interactions.

This model eliminates the necessity for businesses to develop custom integrations with numerous other services, fostering a win-win scenario where the agent serves as a connective layer, simplifying complex workflows for the user.

Users are increasingly rejecting the notion of managing separate agents for every company, such as Amazon, Etsy, or Expedia. The preference is for a single, centralized personal assistant that can handle all interactions, rather than navigating between multiple service-specific agents.

This reflects a clear distinction between 'my agent' and 'their agent,' where users want their assistant to act on their behalf across the web, consolidating fragmented digital experiences into a unified control point.

Google is actively developing WebMCP, a standard focused on 'co-browsing' and standardizing agent-website interactions. Instead of agents relying on screenshots to interpret website content, WebMCP allows websites to directly expose JavaScript tools that agents can utilize.

While current implementations, such as Gemini in Chrome, offer some utility, they are often constrained by dashboard-heavy interfaces. WebMCP aims to create a more direct and efficient communication channel between agents and web services, moving beyond mere visual interpretation.

AI assistants are rapidly becoming the primary entry point to the web, fundamentally altering how users access online services. This shift is evident across the industry, with major technology companies striving to make their applications the 'everything app,' centralizing various functionalities.

Tools like Jira are now frequently accessed directly through integrated development environments (IDEs) and AI assistants, bypassing traditional browser-based interaction. This trend suggests that the 'long tail' of less-frequently visited websites will likely experience reduced traffic as users increasingly rely on personal assistants for their digital needs.

For instance, a 9-year-old child now often defaults to ChatGPT for search queries instead of traditional search engines like Google.

Similarly, an older demographic is observed to limit their daily application usage primarily to essential tools like WhatsApp, the camera, and ChatGPT. This behavior indicates a broader trend of reduced engagement with a multitude of apps and websites, signaling a preference for consolidated, AI-driven experiences.

While agents can interact with websites without full UI dependence for tasks like organizing emails, critical decisions, such as booking a honeymoon, still necessitate a 'last mile' of human interaction through a visual interface.

This duality means that web interfaces are increasingly bifurcating into distinct experiences: one optimized for agent automation and another for human oversight. It is projected that 90% of future online interactions will be mediated by agents, requiring websites to be 'agent-ready' to facilitate this shift.

Consequently, developers must design websites that can cater to both autonomous agent interactions and the nuanced, UI-dependent needs of human users.

A development team switched from Mixpanel to PostHog because Claude, an AI agent, recommended the latter due to its superior integration.

This decision highlights how agents evaluate products based on their ease of interaction without requiring a browser, effectively viewing browser-based interaction as a negative 'memory' for future tasks. This shift suggests that brand trust built over decades can be secondary to an agent's ability to seamlessly utilize a service.

Aura, an agentic web research laboratory, has developed readiness benchmarks to evaluate how effectively a website serves AI agents. This free tool allows companies to assess their websites against agentic protocols, providing insights into their compatibility with AI interactions.

The platform includes a public leaderboard that ranks companies based on their agent-readiness, and feedback is generated directly from agents attempting to interact with the sites, offering a practical measure of their usability in the agentic web.

Aura has launched 'Aura's Journey,' a free tool available at journey.ora.ai, which provides real-time visualization of how agents navigate and interact with websites. Users can specify any website, intent, and agent to observe the agent's path as it attempts to achieve a task.

This tool allows for testing how various agents interpret user intent and search for information, with tens of thousands of such tests already conducted to map and understand agentic behavior across the web.

By running multiple agent harnesses simultaneously, such as Cloud Code, Haiku, Eve, and ChatGPT, Aura's Journey visualizes the distinct navigation paths taken by different agents to achieve the same intent. This demonstrates the varied strategies and efficiencies of each AI.

For instance, ChatGPT consistently showed more effective result discovery compared to other agents in these tests. The tool provides a clear visual representation of each agent's 'journey' across a website, highlighting differences in their interaction patterns and success rates.

Historically, the web relied on search engines, mobile on app stores, and social media on feeds for content discovery.

However, these conventional discovery mechanisms are insufficient for agents that require direct access to MCPs, AI resources, and open API.json files. The challenge lies in creating a scalable and curated system for agents to find and utilize relevant web services, as centralized or closed registries lack the necessary flexibility and breadth.

Key among these is AI catalog.json, a collaborative standard developed by Anthropic, OpenAI, Google, and MCP, which standardizes how websites present themselves to agents.

Additionally, the Agentic Resource Discovery (ARD) standard defines how directories expose resources to agents, and A2A (Agent-to-Agent) standardizes website-to-agent communication. These initiatives collectively aim to establish a unified and machine-readable framework for agents to discover and interact with online services.

Aura Directory is actively implementing these standards by scanning domains and generating machine-readable AI catalog.json files. This process, demonstrated with monday.com, allows agents to query specific MCP and API server configurations directly.

The goal is to ensure full Agentic Resource Discovery (ARD) compliance, facilitating interoperability across the agentic web. By exposing these structured JSON files, Aura is helping to build a web directory that agents can efficiently navigate and utilize to discover services.

This approach focuses on preparing current web infrastructure for the increasing presence of AI agents. Agents are already actively navigating the web, underscoring the urgency for standardized protocols and infrastructure that can support their interactions efficiently and effectively.