# Save — Convert Any Webpage to Clean Markdown (Full Reference)

> Save is a Chrome extension that converts any webpage to clean, structured Markdown with one click. It uses AI-powered content extraction (Google Gemini) with 300+ site-specific prompts to produce well-formatted Markdown optimized for note-taking apps (Obsidian, Notion) and AI workflows (ChatGPT, Claude, Gemini). Built for developers, writers, researchers, and AI practitioners.

This is the full reference document for LLMs and AI agents. A shorter overview is available at /llms.txt.

--------------------------------------------------------------------------------

## 1. What Save is

Save is a three-part product:

1. **Save (Chrome extension)** — captures the current tab and converts it to Markdown via an AI backend. Works on Chrome, Edge, Brave, and Arc (Manifest V3).
2. **Save Vault (macOS menu bar app)** — receives Markdown from the extension via localhost:7421, writes it into `~/Documents/Save Vault/<knowledge-base>/`, and exposes those files to AI clients via MCP.
3. **Save MCP** — an MCP server built into Save Vault. Launched with `--mcp`, it exposes the user's saved pages to Claude Desktop, Claude Code, and any MCP-compatible client.

Together they form a "save the web, read it with AI" loop:
- Save pages in Chrome with one click.
- They land as Markdown files in a local folder that doubles as an Obsidian vault.
- Claude reads them via MCP and grounds answers in the user's own knowledge.

## 2. Website structure

- `/` — homepage (this project's marketing site), available as HTML and as Markdown via Accept: text/markdown or /index.md
- `/pricing` — pricing page
- `/mac` — Save Vault download page for macOS
- `/blog/*` — blog posts (English) with localized variants under `/<locale>/blog/*`
- `/privacy`, `/terms` — legal
- `/welcome` — post-install welcome page (5 locales)
- `/sitemap-index.xml` — sitemap (legacy `/sitemap.xml` redirects here)
- `/robots.txt` — crawler rules (permissive for AI bots)
- `/llms.txt` — short agent-readable overview
- `/llms-full.txt` — this document
- `/.well-known/mcp/server-card.json` — MCP Server Card for Save MCP
- `/.well-known/agent-card.json` — A2A Agent Card (minimal)
- `/.well-known/agent-skills/index.json` — index of skills Save exposes
- `/.well-known/api-catalog` — RFC 9727 link set for Save's HTTP APIs

Locales supported on marketing pages: en (default), fr, es, de, ja, it, ko, pt-br, zh-tw, ru, pl, vi.

## 3. What Save does (extension behaviour)

### 3.1 Removed from output
- Advertisements and tracking scripts
- Navigation menus, sidebars, site chrome
- Cookie and consent banners, pop-ups, modal overlays
- Boilerplate footer links and duplicated navigation
- Inline JavaScript and style noise

### 3.2 Preserved in output
- Headings (H1–H6), preserving document hierarchy
- Ordered and unordered lists, including nested lists
- Inline and block code, with language hints when detectable
- Tables (converted to GitHub-flavored Markdown)
- Links and anchor references
- Image references with alt text and captions
- Blockquotes and emphasis

### 3.3 Capture pipeline
1. **Live DOM preprocessing** (non-destructive): `promoteNoscriptImages()` and `flattenShadowDOM()` enrich the DOM without removing elements.
2. **DOM extraction**: scoring-based article finder (Readability-style) plus clone preprocessing — `normalizeCodeBlocks`, `deduplicateImages`, code-highlighter-aware selectors (Prism, Highlight.js, Shiki).
3. **Screenshot fallback**: when extracted text is empty, dominated by consent banners, or has no real prose, Save captures a full-page screenshot (port of GoFullPage's `page.js`) and sends it to Gemini for visual extraction.
4. **Site-specific prompt**: a router maps the URL to one of 20 site categories (news, e-commerce, social, video, code, docs, recipe, jobs, travel, reference, …). The matching prompt is concatenated with the base template before the AI call.
5. **Post-processing**: the AI response is validated, trimmed, and formatted. Headings are re-anchored and relative URLs are absolutized.

### 3.4 Specialized handlers
- **YouTube videos**: transcript is extracted from the DOM (four fallback methods), summarized by the AI, and returned as a Markdown document with overview, key points, and timestamped sections.
- **Twitter/X threads**: full thread is captured in order with attribution and timestamps.
- **Instagram Reels**: audio is transcribed by the AI and returned as Markdown with caption and oEmbed context.
- **TikTok**: caption, hashtags, and metadata are extracted from the post DOM.

## 4. Supported websites (300+ site rules)

Save ships with AI prompts tuned for categories including:

- **Social media**: Twitter/X threads, Reddit posts and comment trees, LinkedIn articles and posts, Hacker News stories and threads, Mastodon posts.
- **Video**: YouTube videos (AI-summarized transcripts), Vimeo, TED talks, conference recordings.
- **E-commerce**: Amazon product pages (structured specs), eBay listings, Etsy items, Shopify stores, product reviews.
- **Code and docs**: GitHub READMEs and issues, GitLab, Bitbucket, Stack Overflow Q&A, MDN, Python docs, Rust docs, TypeScript handbook, Go doc, Kubernetes docs.
- **Publishing**: Medium, Substack, Ghost, WordPress, personal blogs.
- **News**: NYT, Washington Post, The Guardian, BBC, Reuters, AP, The Atlantic, New Yorker, Wired, FT, Bloomberg, Le Monde, Le Figaro, Les Echos, Süddeutsche Zeitung, Der Spiegel.
- **Reference**: Wikipedia (article, talk, history), Wiktionary, arXiv, Google Scholar results.
- **Recipes**: AllRecipes, Serious Eats, Food Network, NYT Cooking, personal recipe blogs (ingredients/steps normalized).
- **Travel**: Booking.com, Airbnb, TripAdvisor, Hotels.com, expedia, kayak, google travel.
- **Jobs**: LinkedIn jobs, Indeed, Glassdoor, Lever, Greenhouse, Workable.
- **Any other webpage**: a generic extraction prompt handles sites without a specific rule.

## 5. Output quality

- Markdown is valid CommonMark + GFM extensions (tables, strikethrough, task lists).
- Token efficiency: clean Markdown typically uses ~3–5× fewer tokens than the original HTML when consumed by an LLM.
- Deterministic structure across repeated saves of the same page, up to DOM mutations.
- Image references are kept as URLs (no binary embedding) to keep files small.
- Code blocks carry a language hint whenever the source page exposes one.

## 6. Save Vault (macOS app)

Save Vault is a SwiftUI menu bar application:

- **Bundle ID:** md.save.Save
- **Storage root:** `~/Documents/Save Vault/`
- **Default knowledge base:** `General` (created on first launch)
- **HTTP listener:** `localhost:7421` (for the Chrome extension)
- **MCP mode:** launched with `--mcp`, speaks JSON-RPC over stdio

### 6.1 HTTP API (localhost:7421)

| Method | Path | Purpose |
|--------|------|---------|
| GET | /ping | Health check |
| GET | /kbs | List knowledge bases |
| POST | /kbs | Create a knowledge base |
| POST | /kbs/:name/files | Save a markdown file into a KB |
| GET | /search?q=query | Full-text search across all KBs |

### 6.2 MCP tools exposed to Claude

| Tool | Purpose |
|------|---------|
| list_knowledge_bases | List KBs with file counts and last-modified timestamps |
| list_files | List files in a given KB |
| read_file | Read a file by KB + filename |
| search | Full-text search across the whole vault |

Enable the MCP by toggling "Connect to Claude" inside Save Vault. Disable it to return Claude to a stock generalist.

## 7. Pricing and limits

- **Free:** 3 saves per month. Tracked in chrome.storage; resets monthly.
- **Plus:** Unlimited saves. $5.99/month or $34.99/year. A 500 saves/month fair-use policy applies for abuse protection. Billed via Stripe.
- **Chrome Web Store:** https://chromewebstore.google.com/detail/save-%E2%80%94-web-to-markdown/mamnlljnkigkhppbjhmpdeocobcbobdp

## 8. Privacy

- Captured page content is sent to Save's Cloudflare Workers API, then to the AI provider (Google Gemini primary, Anthropic Claude fallback) for extraction.
- Page content is not stored server-side beyond the duration of a request.
- PostHog is used for product analytics on the extension and the marketing site. Stripe handles payments.
- Local saves via Save Vault never leave the user's machine.

See: https://www.minibase.md/privacy

## 9. Terms

See: https://www.minibase.md/terms

## 10. Brand

- **Name:** Save
- **Visual mark:** a bookmark (retired the acorn mark in March 2026; the bookmark is the canonical mark across extension, website, favicon, and macOS app).
- **Primary color:** #F4511E (Save Orange)
- **Wordmark:** "Save"
- **Button style:** pill-shaped, 9999px border radius
- **Spacing base:** 4px

## 11. Contact

- **Email:** hello@minibase.md
- **Web:** https://www.minibase.md

## 12. Machine-readable discovery

- robots.txt: https://www.minibase.md/robots.txt
- Sitemap: https://www.minibase.md/sitemap-index.xml
- LLMs.txt (short): https://www.minibase.md/llms.txt
- LLMs.txt (full): https://www.minibase.md/llms-full.txt
- MCP Server Card: https://www.minibase.md/.well-known/mcp/server-card.json
- A2A Agent Card: https://www.minibase.md/.well-known/agent-card.json
- Agent Skills Index: https://www.minibase.md/.well-known/agent-skills/index.json
- API Catalog: https://www.minibase.md/.well-known/api-catalog
- Markdown alternate: https://www.minibase.md/index.md (also served on `/` with Accept: text/markdown)

---

# Save API (for developers and AI agents)

Hosted HTTP API that turns any URL into clean Markdown — the same engine as the extension, callable from code. Built for AI agents, RAG pipelines, and scrapers.

- Endpoint: POST https://api.savemarkdown.co/v1/convert
- Get a free key (no human in the loop): POST https://api.savemarkdown.co/v1/signup with {"email":"you@example.com"} → returns a sk_test_ key, 100 conversions/month free, no card. Then use it as the Bearer token.
- Auth: Authorization: Bearer sk_live_… (server-side only; never ship a key in a browser)
- Body: {"url": "https://…", "template?": "clean|summary|outline|…", "render?": "auto|never|always", "fresh?": false}
- Response: {"markdown": "…", "meta": {"tier": 1, "cached": false, "domain": "…", "template": "markdown"}, "usage": {"credits": 1}}
- Tiered engine: cheap server-side fetch → headless render only when a page is a JS shell → boilerplate stripped. meta.tier tells you which rate applied (1 = fetch/extract, 2 = render, "adapter" = platform channel).
- Pricing: Markdown $2 / 1,000 pages · Rendered (JS) $8 / 1,000 pages · AI-formatted templates $3/$25 per 1M input/output tokens. Free trial 100 pages. Pay as you go.
- Walled gardens: YouTube via official transcript channel; X/Instagram/TikTok best-effort; no ghost accounts or login-scraping.
- Agent spec (markdown): https://www.minibase.md/api.md
- Human docs: https://www.minibase.md/api
- Request a key: hello@minibase.md
- Discovery: agent-skills id save.api.url-to-markdown; /.well-known/api-catalog.json