Summary

It starts working with zero configuration, and every key you add later buys one more class of page it can reach.

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.

What it is

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.

What you get

fetch returns a URL as clean markdown, with dedicated handling for tweets, YouTube videos including transcripts where available, arXiv papers, PDFs, Wikipedia articles, and GitHub repos, files, issues, PRs and releases — direct image URLs come back as an image block for vision
fetch_batch pulls up to 10 URLs in parallel through the same chain, with a smaller per-URL length budget
search across the web with domain filtering, freshness and paging — Brave when a key is set, then a SearXNG instance, then DuckDuckGo as a last resort
research does the search and the fetching in one call, returning the full content of the top results ready to compare
extract pulls named fields out of a page using CSS selectors instead of prose, and can convert every HTML table into arrays of row objects
A tiered fallback chain that keeps going when the first strategy fails — up to fourteen more before it gives up
Alongside the tools: a recent-session resource, and prompts for researching a topic and extracting an article

Requirements

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.

Setup effort

One command — npx -y intercept-mcp

Intercept