Show HN: Spegel, a Terminal Browser That Uses LLMs to Rewrite Webpages

Concept and Use Cases

• Spegel is seen as a clever way to browse the web as text from the terminal, using LLMs as a “user agent” that works for the user rather than site owners.
• People imagine multi-tab workflows: compare multiple news outlets and Wikipedia, then have the tool summarize and reference differences.
• Several see strong potential for accessibility and screen-reader–like interfaces, or for low-bandwidth / older hardware browsing.
• Others want it as a proxy service: strip cruft and ads server-side, then serve clean text to any browser.

Interaction Model and Extensions

• Suggested features: multiple views per page (original vs fact-checked), post request handling, scripting, prompt flags like -p "extract only the product reviews".
• Integration ideas: Emacs (eww), Lynx (stdin), Chrome extensions, MCP tools, and using it as a backend for agents.
• Some propose caching per-URL/per-prompt outputs or sharing “lenses” so pages aren’t reprocessed every visit.

Technical Approaches and Performance

• Proposals to reduce tokens and cost: run Firefox Reader / Readability first, or use deterministic HTML→Markdown tools (pandoc, semantic markdown libraries, schema.org Recipe).
• Others suggest headless Puppeteer/Selenium to render JS-heavy SPAs, then feed the DOM to an LLM or simpler extractor.
• Many argue a small local model would be more appropriate than a big cloud LLM for routine conversion.

Reliability, Hallucinations, and Determinism

• Major concern: non-determinism and silent content changes.
• The recipe demo became a focal point: users documented that ingredient amounts and items were subtly but significantly altered.
• The author later confirmed truncation caused the model to hallucinate a known recipe from training, using this as proof that models can’t be fully trusted.
• Critics say this undermines accessibility (which requires predictability) and makes LLM-based “rewriting” fundamentally suspect, especially for anything safety- or fact-critical.

Web Ecosystem, SEO, and Ads

• Some praise LLM filtering as a way to fight SEO junk and overlong recipe blogs; others note deterministic solutions and structured data already exist but are underused.
• Discussion of recipe-site economics: long posts and ads driven by SEO, with LLM-based “reader” layers potentially disrupting this model.
• Speculation about future “LLM ad blockers” vs “SEO for AI,” and worries about reinforcing personal bubbles or “memetic firewalls.”