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 for manual clipboard capture or an optional local MCP workflow. Manual mode Savage Scraper works out of the box with no setup required. Open any normal web page and click the Savage Scraper toolbar action once. Savage Scraper captures the rendered active page, simplifies its useful structure, and copies the result directly to the local clipboard. Manual mode is the default. It does not require configuring Savage MCP, granting broad website access, or setting up any external service. Optional MCP mode MCP mode is disabled by default. If the user installs the separate local Savage MCP server and explicitly enables MCP website access in Savage Scraper, an MCP client can read selected HTTP/HTTPS pages through the user's existing Chrome profile and browser session. The local Savage MCP configuration is the source of truth for MCP website access. The user configures allowed_hosts and optional allowed_paths in Savage MCP's local config file. After the localhost bridge has been mutually authenticated, Savage MCP sends the normalized host/path policy to Savage Scraper. Savage Scraper keeps that policy in memory and independently enforces it before opening or scraping pages. The host/path policy is not stored as a separate whitelist in Savage Scraper's Chrome settings. allowed_hosts is the outer access boundary. Optional allowed_paths rules can further restrict an allowed host to specific paths. A missing allowed_paths entry or an empty array allows all paths on that already-allowed host. A non-empty entry can restrict access to exact paths or recursive path prefixes. For example, access to github.com can be limited to a single repository instead of granting access to every GitHub repository. The local bridge connects only to Savage MCP on 127.0.0.1 and uses mutual HMAC authentication with a shared local token. The token itself is not transmitted over the WebSocket. The MCP browser surface is intentionally limited to opening allowed URLs, scraping one dedicated agent tab, and reporting status. It does not expose generic clicking, typing, form submission, arbitrary JavaScript execution, or unrestricted browser control. Before an MCP scrape, Savage Scraper performs a bounded page-settling and main-document scrolling pass to give dynamically rendered and lazy-loaded content time to appear. It restores the original main-page scroll position before extraction. Nested scroll containers are not traversed. Savage Scraper has no developer-operated backend and contains no analytics, advertising, telemetry, or remotely hosted executable code. In manual mode, the generated result is written directly to the user's local clipboard. In MCP mode, the generated result is returned over the authenticated loopback connection to the user's locally running Savage MCP process, which returns it to the user's chosen MCP host. Open source: Savage Scraper: https://github.com/dominikduda/savage_scraper Savage MCP: https://github.com/dominikduda/savage_mcp Privacy policy: https://github.com/dominikduda/savage_scraper/blob/main/PRIVACY.md