LLM browser generates webpages instead of fetching them

The LLM browser is a desktop application that inverts the traditional web browsing model. When you type an address (for example, google.com), instead of making an HTTP request to a remote server, the app sends a prompt to an LLM endpoint you have configured, and the LLM generates the webpage content as if that site existed. All standard browser features work: you can click links, use search boxes, open multiple tabs, and navigate freely. Each new page is generated with knowledge of the previous page's URL, referrer, and browsing history, allowing the LLM to maintain narrative coherence across an entire browsing session.

To use the app, you open Settings and configure your LLM provider. The available options include OpenAI (with default endpoint https://api.openai.com/v1 and example model gpt-4o) and Anthropic (endpoint https://api.anthropic.com, example model claude-sonnet-5). You then enter your API key, which is stored locally in the app and never exposed in the UI again. Once configured, you can begin typing addresses and exploring the LLM-generated web.

A distinctive feature is the "Universe rules" setting, accessible via Settings → Universe rules. By default, the app generates sites that correspond to the real world. However, you can rewrite these rules to change the entire premise of the generated universe. The documentation provides an example: a universe where the Soviet Union still exists and humanity grows stronger and happier. Once you save custom rules and navigate to any site, the LLM will generate all subsequent pages — including sites, brands, news articles, and search results — consistent with your new universe's premises. Clearing the field restores the default real-world rules.

The project includes a test suite (npm run test:e2e) powered by Playwright and a mock LLM, so developers can run tests without requiring actual API keys or network access. This engineering foundation suggests the creator intended the tool for exploration and experimentation with how LLMs can sustain large-scale, rule-governed simulated environments.

This project reimagines web browsing by replacing HTTP requests with LLM-driven generation. Rather than the browser fetching real webpages, it instructs an LLM to produce coherent HTML/content for each URL and interaction, maintaining session state (referrer, history) so that navigation feels natural and the simulated web environment remains internally consistent across multiple clicks and searches. The architecture is minimal — only the LLM endpoint receives requests; the app itself stores the API key locally. This approach sidesteps traditional server-client networking entirely, which may appeal to developers exploring how LLMs can simulate complex, rule-based environments. The "universe rules" feature is particularly notable: by allowing users to redefine the foundational premises (e.g., alternate history scenarios), it demonstrates how LLM-generated content can scale from a single page to an entire coherent world-model that respects user-specified constraints across all subsequent interactions.

No comments yet. Be the first to share your thoughts!