OpenBento

A living intelligence dashboard where humans and AI agents organize the web together.

Comment

Inspiration

The web is still mostly designed around humans opening tabs, searching pages, and manually organizing what they find. AI agents can answer questions about the web, but they usually operate in a separate chat experience rather than inside the same persistent workspace as the user.

OpenBento explores a different model: what if humans and agents could work on the same web-native intelligence dashboard?

What it does

OpenBento is a persistent dashboard for following topics as they evolve.

Each Canvas contains one live Dashboard where users can combine persistent WatchBots, web/news Cards, X posts with rich media and engagement data, playable YouTube Cards, notes, structured Columns, financial Cards, and AI assistance.

A WatchBot owns a dedicated Column. New discoveries appear at the top of that stream instead of endlessly expanding the workspace. When something matters, the user can drag that Card out of the live feed and place it directly onto the Dashboard, preserving it while the WatchBot continues monitoring.

The result is somewhere between a live research desk, TweetDeck, an intelligence dashboard, and an AI workspace.

Why WebMCP is a strong fit

OpenBento is fundamentally about shared state between humans and agents.

Without WebMCP, an AI assistant can describe what the user should do, but it cannot naturally operate the same web application the user is already working in. With WebMCP, OpenBento exposes meaningful application actions directly to compatible agents.

An agent can work with the same persistent Canvases, Cards, WatchBots, and Dashboard state that the human user sees. The agent is not controlling the application through screenshots or brittle coordinate-based browser automation. It interacts through explicit tools published by the website itself.

That makes agent actions more reliable, understandable, and aligned with the application's own permission and domain model.

What humans and agents can do together

A person can ask an agent to create or modify parts of the workspace while continuing to interact with the same Dashboard manually.

For example:

  1. A user asks an agent to create a WatchBot for Apple news.
  2. OpenBento creates a persistent WatchBot and its dedicated live Column.
  3. The WatchBot begins collecting relevant web, X, and YouTube developments.
  4. The user pulls an important Card out of that stream and pins it elsewhere on the Dashboard.
  5. A WebMCP-capable agent can later create Cards, move content, update the Canvas, or manage the WatchBot.
  6. Both the human and the agent continue working against the same persistent state.

This turns the agent from an external chatbot into a participant in the user's workspace.

How WebMCP is implemented

OpenBento registers WebMCP tools through document.modelContext.registerTool.

The WebMCP layer maps a safe set of snake_case tools onto the same authoritative domain action system used by the application's own UI. Current tools cover retrieving Canvas state, creating and switching Canvases, creating/updating/moving/resizing Cards, managing viewport state, fullscreen Dashboard presentation, creating and managing WatchBots, and reading WatchBot status.

OpenBento currently exposes 15 WebMCP tools.

The WebMCP implementation does not maintain a separate privileged mutation layer. Authentication, ownership checks, validation, and persistence pass through the same shared domain executor used by the human UI.

Architecture

OpenBento uses Next.js 16, React, TypeScript, Supabase PostgreSQL/Auth, Railway, XYFlow, OpenAI, X API, YouTube Data API, Sentry, and Resend. The application runs as a web service plus dedicated background workers for different monitoring providers.

Challenges

One of the biggest challenges was making background AI activity coexist with a persistent spatial interface. An early version placed every newly discovered Card directly onto an infinite Canvas. That quickly became unusable.

We redesigned OpenBento around a finite Dashboard with dedicated streaming Columns. WatchBots continuously update their Columns while users can pull important information out into the persistent Dashboard.

We also had to make WebMCP a genuine part of the application's architecture instead of a demo-only wrapper. WebMCP actions use the same domain action catalog, authentication, validation, and persistence layer as the human UI.

What we learned

The most interesting lesson was that agent-friendly websites benefit from exposing intent-level operations, not reproducing mouse and keyboard interaction. A tool like create_watchbot or move_card communicates much more useful application semantics than asking an agent to infer coordinates and manipulate the UI visually.

WebMCP makes that model native to the website.

What's next

OpenBento can expand into a broader personal intelligence environment with richer market Cards, more social/news providers, AI-generated layouts, configurable Dashboard templates, team collaboration, and additional WebMCP actions.

AI organizes the story. Sources remain the story.

Built With

Share this project:

Updates