Herbarium Saves AI-Generated HTML for Spaced Review

Overview

A new desktop application called Herbarium is addressing a growing pain point in the AI era: the ephemeral nature of generated content. As users increasingly rely on large language models to produce HTML-based explanations, documentation, and interactive tutorials, the risk of losing these valuable resources after a browser session ends has become significant. Herbarium offers a local-first solution that treats AI-generated pages as durable, reviewable assets rather than disposable web content.

The application functions as a dedicated reader for HTML pages produced by AI tools. Its core value proposition lies in its ability to preserve the interactivity of these pages while organizing them into a searchable, schedulable vault. By storing data as plain files, Herbarium ensures that users maintain full ownership of their knowledge base, independent of any proprietary platform or subscription service.

How It Works

Herbarium operates on a simple yet robust file structure designed for longevity and compatibility. Each saved page consists of a standard .html file paired with a .json metadata file. This metadata includes the page title, tags, user notes, and review scheduling data. The use of plain text formats means the vault can be backed up, synchronized via cloud storage, or edited with any preferred tool outside the application.

The application supports multiple methods for importing content. Users can drag and drop files directly into the window, use a standard file picker, or paste HTML code directly from their clipboard. Once imported, pages are stored in a local directory that acts as the user’s personal knowledge vault. Because the files are native to the operating system, changes made externally are detected and reflected within the app upon return.

Your pages stay plain .html files in a folder you own.

Safety and Sandboxing

Given that HTML is an executable format capable of running arbitrary code, Herbarium prioritizes security through strict sandboxing. Every imported page is treated as untrusted until verified by the user. The application isolates the rendered content in an environment that cannot access the user’s local files, interact with other applications, or navigate away from the sandboxed viewer without explicit confirmation.

Links embedded within the generated pages do not open within the app. Instead, they require user consent before launching in an external web browser. This design prevents malicious or accidental navigation to unintended destinations while allowing users to explore linked resources safely.

Additionally, Herbarium includes a per-page network access toggle. When enabled, pages can fetch scripts, fonts from well-known CDNs, and images from external sources. However, this feature carries privacy implications, as some scripts may contain analytics or tracking pixels that notify the original author when the page is opened. Users are advised to keep network access disabled for untrusted or suspiciously generated content.

Review and Organization

One of Herbarium’s distinguishing features is its integration of spaced repetition for content review. Unlike static PDFs or bookmarked URLs, AI-generated HTML pages often contain complex explanations or code snippets that benefit from periodic reinforcement. Herbarium allows users to schedule pages for review at intervals they define, ensuring that important information is revisited when it is most needed for retention.

The application provides a keyboard-driven review session interface, making it efficient for power users who wish to process multiple pages in quick succession. Reminders can be configured to prompt users before scheduled reviews, helping to maintain a consistent learning routine.

Organization is handled through a flexible system of folders, tags, and notes. Users can move, edit, or categorize multiple pages simultaneously. A global search function allows for quick retrieval of content based on titles, tags, notes, or the full text of the pages themselves. A command palette (accessible via Ctrl/⌘ K) provides shortcuts for nearly every action, streamlining the workflow for those familiar with modern IDEs or terminal-based tools.

Integration with AI Agents

Herbarium is not merely a passive archive; it is designed to be an active component in AI-assisted workflows. The application includes an MCP (Model Context Protocol) server, enabling AI agents to interact with the vault directly. This allows users to ask their AI assistants to save, search, organize, and schedule pages without manual intervention.

For example, a user working in Claude Code can add the Herbarium MCP server and then issue commands such as, “Write a short HTML explainer of Cargo workspaces and save it to Herbarium under rust.” The agent will generate the content, format it as HTML, and store it in the user’s specified folder, complete with appropriate tags and metadata. This integration transforms Herbarium from a simple reader into a persistent memory layer for AI-driven development and research tasks.

Installation and Availability

Herbarium is built using Tauri 2, combining Rust for performance and security with Svelte 5 for the user interface. It is available as a desktop application for Linux, macOS, and Windows.

On Linux, users can install the application via a shell script that automates the setup process for both Debian/Ubuntu systems (producing a .deb package) and other distributions (producing an AppImage). macOS and Windows users can download pre-built binaries from the project’s release page. Due to the current lack of code signing, users may need to adjust their system settings to allow the application to run on first launch.

The project is open-source and licensed under the MIT License. Developers interested in contributing or building from source can find instructions in the repository, which requires Node.js, pnpm, Rust, and Tauri prerequisites.