What it does
codearia-sieve fetches a web page and converts it before a model reads it. Dates are normalised to ISO format, numbers are recorded as facts only when they carry a unit, and the remaining text is split into chunks sized in tokens and characters. Each chunk links to an anchor on the source page. Menus, footers, banners and similar material are removed, and with tracing enabled the output lists what was removed and why. The same HTML always produces the same JSON, and no model runs unless one is connected.
It suits developers who pay per token for fetched pages and want a smaller input for the model. It also suits anyone running small classifier or ranker models, and anyone who needs each claim traced to its source. The publisher's sample of random pages had a median of 53,718 input tokens and 1,106 output tokens. The sieve_page tool returns a structured result, and in its default summary mode the chunks carry sizes, anchors and headings but no text. The sieve_chunk tool returns the text of one chunk without refetching the page.
It runs as an MCP server, listed in the official MCP Registry under its registry name, so compatible clients can install it by that name. It fetches pages over plain HTTP, checks robots.txt first and uses no browser. Problems such as blocked requests, paywalls and HTTP errors come back as named warnings rather than exceptions.
Limitations and requirements
Sites whose robots.txt disallows crawling are not fetched. Bot challenges and refusals (403, 405, 429) are reported as blocked. Paywalled articles return only the teaser. Content that JavaScript would fill in may be empty, because no browser is used. Pages with more than 500 facts are capped at 500 listed. Default chunk budgets are 20,000 tokens and 50,000 characters.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.AntonG87/codearia-sieve
Web pages into decision-ready state: dates, numbers with units, budgeted chunks. No model, no key.
v0.2.2As shown on registry.modelcontextprotocol.io when we last read it