Savage Scraper

Scrapes rendered pages into simplified LLM-friendly HTML for clipboard or optional local MCP workflows.

Savage Scraper converts rendered web pages into simplified, LLM-friendly HTML. It supports two workflows: • Manual mode — click the Savage Scraper toolbar icon to scrape the active page and copy the result directly to your clipboard. • Optional MCP mode — connect Savage Scraper to the separately installed, locally running savage_mcp server so an MCP-compatible AI client can open and scrape explicitly whitelisted websites in your normal Chrome browser. Manual mode is the default and does not require MCP setup. ## Features • One-click scraping of the active page • Copies manual scrape results directly to your clipboard • Extracts rendered page content instead of raw source HTML • Preserves useful semantic HTML structure • Keeps headings, paragraphs, lists, tables, links, forms, and other meaningful elements • Preserves useful attributes such as links, selected ARIA labels, IDs, form state, and image alt text • Filters framework-generated and utility-style CSS classes • Supports compact and pretty-formatted HTML output • Optional inclusion of hidden and collapsed page content • Supports extraction of xterm.js terminal content when accessible • Includes useful page metadata such as URL, title, capture time, and viewport size • Optional local MCP integration for automated scraping of user-whitelisted websites • Reuses a dedicated Chrome tab for MCP operations • Can automatically close the MCP tab after inactivity or immediately after a successful scrape • Stores extension preferences locally • No account required • No analytics • No advertising • No developer-operated backend service • No remotely hosted executable code ## Why use Savage Scraper? Copying a web page with Ctrl+A / Ctrl+C often produces incomplete or awkward results. Copying raw page source has the opposite problem: it contains scripts, CSS, framework internals, generated class names, metadata, and large amounts of implementation detail that are usually irrelevant when working with an LLM. Savage Scraper sits between those two approaches. It walks the rendered page and creates a smaller representation intended to preserve useful information while removing unnecessary implementation detail. The output is intentionally not a 1:1 copy of the original HTML. It is designed to be compact, understandable, and practical for LLMs, technical notes, bug reports, documentation workflows, and other text-based uses. ## Manual mode Open the page you want to capture and click the Savage Scraper toolbar icon. Savage Scraper runs its packaged scraper on the active page and copies the generated simplified HTML to your local clipboard. Manual mode uses Chrome's temporary active-tab access and does not require permanent website access. ## Optional MCP integration Savage Scraper can optionally connect to the open-source savage_mcp project: https://github.com/dominikduda/savage_mcp savage_mcp is a local Model Context Protocol server that allows MCP-compatible clients to request scraping through Savage Scraper. MCP mode is disabled by default. To use it, the user must: • install and run savage_mcp on the same computer; • configure an explicit allowed_hosts whitelist; • configure the matching local bridge token and port in Savage Scraper; and • explicitly approve Chrome's optional HTTP/HTTPS website-access permission. The extension communicates with savage_mcp through an authenticated WebSocket connection bound to the local loopback interface at 127.0.0.1. Actual MCP navigation and scraping are restricted to the allowed_hosts whitelist configured by the user. The whitelist is enforced by both savage_mcp and Savage Scraper. Savage Scraper does not expose arbitrary JavaScript execution, generic clicking, typing, or unrestricted browser control through MCP. ## MCP browser behavior MCP mode uses one dedicated Chrome tab. When an MCP client requests an allowed URL, Savage Scraper can create or reuse that tab, load the page, perform a limited main-page scrolling pass to trigger common lazy-loaded content, restore the page position, and return the simplified HTML to savage_mcp. The dedicated tab normally closes after a configurable inactivity timeout. Users can also enable close_after_scrape in savage_mcp. When enabled, the dedicated MCP tab closes immediately after a successful scrape instead of remaining available for reuse. ## Typical uses Savage Scraper can be useful for: • Pasting page content into ChatGPT or another LLM • Providing webpage context to an MCP-compatible AI client • Capturing documentation or technical pages • Copying structured tables and lists • Preparing webpage context for bug reports • Capturing rendered application state • Saving simplified page content into notes • Extracting visible text together with useful HTML structure • Copying terminal output from supported xterm.js interfaces ## Hidden content By default, Savage Scraper attempts to include only rendered and visible page content. You can enable "Include hidden content" to also include hidden or collapsed DOM content where accessible. Password input values are never included. Hidden input fields are excluded entirely. ## Form content Savage Scraper may include the current values of ordinary form controls because those values can be part of the page state the user intends to capture. Password fields and hidden input fields are never included. ## xterm.js terminal support Some web-based terminals use xterm.js and render content in ways that normal DOM scraping cannot reliably capture. Savage Scraper attempts several extraction methods, including: • xterm.js buffer access • accessibility-tree extraction • older DOM-based xterm rendering When the actual terminal buffer is accessible, scrollback content may also be included. ## Output format Each scrape includes basic metadata followed by the simplified page representation. Typical metadata includes: SCRAPED_BY PAGE_URL PAGE_TITLE CAPTURED_AT VIEWPORT The resulting HTML can be generated in either compact or pretty-formatted form. Compact mode minimizes output size. Pretty mode adds indentation and line breaks to make the result easier to inspect manually. ## Privacy Savage Scraper does not contain analytics, advertising, telemetry, or a developer-operated backend service. In manual mode, page content is processed locally in Chrome and the generated result is written to the user's local clipboard. In optional MCP mode, the generated result is sent only to the user's locally running savage_mcp process through the localhost bridge. savage_mcp then returns that result to the user's chosen MCP host. The user's MCP host and any model provider configured in that host may subsequently process the scraped content according to the user's configuration and those services' own privacy policies. Savage Scraper does not send scraped page content to the developer. For full details, see the Savage Scraper privacy policy: https://github.com/dominikduda/savage_scraper/blob/main/PRIVACY.md ## Permissions Savage Scraper uses the following required Chrome permissions: • activeTab — provides temporary access to the current page when the user explicitly invokes manual scraping • scripting — runs the packaged Savage Scraper extraction code and MCP lazy-load scrolling code on authorized pages • storage — stores extension preferences, local MCP settings, and session-scoped MCP tab state • clipboardWrite — copies manual scrape results to the user's clipboard • alarms — manages MCP agent-tab cleanup and local bridge reconnection across Manifest V3 service-worker suspension Savage Scraper also declares optional HTTP/HTTPS website access. This optional website permission is not required for normal manual scraping and is not requested automatically when the extension is installed. It is requested only when the user explicitly enables MCP website access in Savage Scraper's extension options. MCP navigation and scraping remain additionally restricted by the user's allowed_hosts whitelist. ## Security The optional MCP bridge: • listens only on 127.0.0.1 • uses mutual challenge-response authentication based on a user-local shared bridge token • does not transmit the bridge token directly • accepts only a fixed set of supported MCP actions • does not execute remotely supplied JavaScript or other executable code • validates the user-configured website whitelist on both sides of the bridge ## Limitations Some browser and page technologies cannot be fully captured. Known limitations include: • Chrome-protected pages such as chrome:// pages cannot be accessed • Chrome Web Store pages cannot be scripted • iframe contents are skipped • SVG contents are skipped • Canvas contents are skipped, except for supported xterm.js extraction paths • Shadow DOM contents are not explicitly traversed • Visibility detection is heuristic • MCP lazy loading scrolls only the main page, not arbitrary nested scroll containers • xterm.js extraction depends on how a website exposes its terminal implementation Savage Scraper is open source. Source code: https://github.com/dominikduda/savage_scraper