MCP SERVERGets the page when a plain fetch returns a 403 or a wall of HTML — clean markdown out, with tweets, YouTube transcripts, PDFs, arXiv and GitHub handled specially.
LABSCO SUMMARY
The fallback ladder is the whole design: a site-specific handler if one exists, then progressively heavier fetchers, so most URLs come back as usable markdown rather than an error. Add nothing up front — reach for a search key when DuckDuckGo throttles you, Cloudflare Browser Run for single-page apps, and FlareSolverr only for the pages that hand you a challenge.
A web-reading server built as a ladder: site-specific handlers first, then a chain of fallback fetchers, each one harder than the last, until something returns the content.
Nothing at all to start: npx -y intercept-mcp, and fetch works with no keys. Everything else is optional and buys a specific capability. BRAVE_API_KEY or a self-hosted SEARXNG_URL make search reliable — the DuckDuckGo fallback is rate-limited and ignores freshness and paging. GITHUB_TOKEN lifts GitHub's unauthenticated 60-requests-per-hour ceiling. CF_API_TOKEN with CF_ACCOUNT_ID renders JavaScript-heavy pages, FLARESOLVERR_URL solves challenge pages at 30-60 seconds each, and WEB_UNLOCKER_URL points at a commercial unlocker as the paid last resort. INTERCEPT_AUTH, a JSON map of domain to headers or cookies, lets it read pages you are signed in to. Caching and proxying are tunable too, including turning the shared agentsweb.org cache off or read-only.
One command — npx -y intercept-mcp
USE CASES
Hand over the URL and let the fallback chain find a strategy that works.
Search and fetch the top results in one step, and read them side by side.
Search first, then batch-fetch the URLs worth reading in parallel.
Fetch the YouTube URL; the transcript comes back as markdown when one is available.
Name the fields with CSS selectors, or ask for every table as rows of objects.
Attach your headers or cookies for that domain, then fetch as normal.
Counted once with only startup setup. Other versions, modes or optional credentials may expose a different list.
extractExtract specific values from a web page as JSON instead of markdown prose. Provide CSS selectors to pull named fields (text or an attribute, first match or all), and/or set tables:true to convert every HTML table to arrays of row objects. Use this when you need particular data (prices, lists, specs, tabular data) rather than the whole page. Honors per-domain auth and proxies.
fetchFetch a URL and return its content as clean markdown. Handles Twitter/X tweets, YouTube videos (with transcripts), arXiv papers, PDFs, Wikipedia articles, and GitHub repos/files/issues/PRs/releases directly. Direct image URLs (png/jpeg/gif/webp) are returned as an image block for vision. Otherwise checks a shared cache, then falls back through a multi-tier chain: Jina Reader, web archives (Wayback, archive.ph, Arquivo.pt), raw fetch, RSS, CrossRef, Semantic Scholar, HackerNews, Reddit, OG meta. Long pages are truncated at maxLength characters — paginate with startIndex. Results are cached for the session; pass noCache to force a live fetch.
fetch_batchFetch up to 10 URLs in parallel and return each as clean markdown. Same handler/fallback chain as the fetch tool, with a smaller per-URL length budget. Use after a search to pull several sources in one call.
researchSearch the web and fetch the content of the top results in one call. Returns the full content of each result as markdown, ready to summarize or compare. Use this instead of separate search + fetch calls when researching a topic.
searchSearch the web and return results. Uses Brave Search API if BRAVE_API_KEY is set, otherwise falls back to SearXNG and then DuckDuckGo. Supports domain filtering (site), freshness, and pagination (page).
BEFORE YOU INSTALL
5 tools when we connected. The environment variables change fetch strategies, not tools: an unconfigured install has a DuckDuckGo-only search and no JS rendering.
CONTEXT COST
Rated by the size of the 5 tool definitions that ride along on every turn, not by how many there are
INSTALLATION
npx -y intercept-mcpclaude mcp add intercept -s user -- npx -y intercept-mcpcodex mcp add intercept -- npx -y intercept-mcp{
"mcpServers": {
"intercept": {
"command": "npx",
"args": [
"-y",
"intercept-mcp"
]
}
}
}MCP servers
levz0r
bhubbb
Crawleo
Khamel83