Show HN: Chimera Browser Lets Web Surfers Swap HTML for Live LLM Rewrites

Chimera is a lightweight desktop browser written in Go that uses WebKitGTK for rendering. Instead of directly showing the fetched page, it first scrapes the target site into structured data. You can then render that data either with an inbuilt HTML template or by handing it to a local LLM that synthesises a bespoke page.

• WebKitGTK UI with URL entry and two rendering modes (scrape-only, LLM composed)
• Structured scraping pipeline that extracts titles, headings, highlighted paragraphs, and outbound links
• Optional hand-off to a local LLM endpoint (e.g. Ollama, llama.cpp HTTP server) for bespoke HTML generation
• Polished GTK interface with glassmorphism-inspired styling, client-side decorations, inline spinner, and navigation-aware controls.
• Graceful fallback to the built-in summary template when the LLM is disabled or fails

Ensure the GTK and WebKit development libraries are installed. On Debian/Ubuntu:

On Fedora:

The module depends on github.com/gotk3/gotk3 and github.com/PuerkitoBio/goquery. Fetch them with:

Note: Network access is required to download the modules. If the environment is sandboxed, allow outbound access temporarily or vendor the required modules.

Vendored GTK bindings: Because upstream gotk3 v0.6.4 is missing an import needed when building with recent Go releases, the repository vendors a patched copy under third_party/gotk3. Remove the replace directive in go.mod if you are using a newer upstream release that already includes the fix.

The UI exposes four primary controls:

• A URL entry box. Press Enter to trigger scraping. Press Enter to trigger scraping; it defaults to the "Scrape Only" flow unless you explicitly click "LLM Compose".
• Scrape Only button to build a summary using the internal template.
• LLM Compose button to call the configured OpenAI-compatible endpoint; the model infers the page theme, preserves every detail, and renders a tailored HTML experience.
• LLM Settings button to edit the endpoint, model, API key, and default behaviour at runtime.
• Click any link inside the rendered page to fetch and render that destination using the current mode.

Set the following environment variables before launching the app:

• CHIMERA_LLM_BASE_URL or CHIMERA_LLM_ENDPOINT: Base URL of an OpenAI-compatible API. The client automatically appends /v1/chat/completions if the path is omitted. Examples:
  • http://localhost:11434 (Ollama with openai compatibility)
  • https://api.openai.com/v1
  • https://openrouter.ai/api
• CHIMERA_LLM_MODEL: Name of the chat completion model (e.g. gpt-4o-mini, mistral-nemo, llama3).
• CHIMERA_LLM_API_KEY: Bearer token for providers that require authentication (OpenAI, OpenRouter, etc.). Leave empty for unauthenticated local endpoints.
• CHIMERA_USE_LLM=1 (optional): Auto-trigger the LLM path when pressing Enter in the URL field.

The request payload matches the OpenAI Chat Completions schema. The system prompt instructs the model to emit a complete HTML document; the scraped data is supplied as a single user message. Responses are expected in the choices[0].message.content field. Chimera persists LLM settings to ~/.config/chimera/settings.json; updates made in the UI become the new default unless overridden by environment variables. If the LLM returns a rate-limit (HTTP 429), Chimera automatically falls back to the scraped view and switches subsequent link navigations to template mode until you manually trigger LLM Compose again. The assistant re-styles the page but must not summarise or drop content; all sections, wording, and links from the scrape are preserved in the generated HTML.

When the endpoint is reachable the LLM Compose button becomes active. Errors from the LLM call are surfaced inside the web view and the app automatically falls back to the template-based rendering.

• Persist browsing history and scraped datasets locally.
• Cache scraped results to avoid repeated downloads when iterating with the LLM.
• Provide configuration UI for toggling automatic LLM usage and model selection at runtime.
• Add tests for the scraper pipeline (mocking responses) and the HTML renderer.