Bintu Web Reader

Fetch one public HTTP or HTTPS URL and return its readable text, sized for an LLM context window. Only HTML/XHTML pages up to 1 MiB are supported (PDF, JSON, images and pages that need JavaScript are not). Article or main content is preferred; navigation, headers, footers, ads and cookie banners are dropped, repeated story blocks are removed (deduplicated_count), links appear as 'label (url)', code is returned in fenced blocks with indentation kept, and tables are pipe-delimited. Modes: scrape (default) returns the text trimmed to max_tokens; chunk splits it into chunks that each fit max_tokens (headings stay with their text, table header rows repeat in continuation chunks, and one indivisible unit such as a very long table row can exceed the budget); summarize returns the leading lines and whole sentences that fit max_tokens (it is not an AI-written summary). Pages are read up to about 200,000 characters (roughly 50,000 tokens): truncated=true with limiting_factor=char_cap means that ceiling was hit, token_budget means max_tokens was. tokens and approx_tokens are approximate estimates (characters divided by 4), not tokenizer counts; tokens_saved_estimate is the source-HTML estimate minus the tokens delivered in this response. low_yield=true means a large page produced very little text, usually because it needs JavaScript or a login; see yield_hint. Results may come from a cache (cache_hit, cache_layer): set cache=false to bypass cache reads and persist=false to skip the persistent write; URLs with credential-like query parameters are never stored. robots.txt is honored. JavaScript rendering (render_js) works only if the operator enables it. Failures return isError=true with a code such as invalid_or_unsafe_url, upstream_access_denied, blocked_by_robots, payload_too_large or unsupported_media_type.