Skip to main content
Hand smry a link and get back clean text, or just the paragraphs that answer your question.

Ask a question instead of reading the page

An agent that reads a whole article to find one sentence burns the context it needs for the actual work. Ask smry the question instead.
what did they change about team pricing in this post? https://harbor.example/blog/pricing-update
search_article · 1 call · returned 3 passages, not the page smry ranked every paragraph in the article against the question and returned the three that answered it. ¶12 is that paragraph’s real position in the piece, so your agent can quote it instead of paraphrasing from memory. The whole page would have cost roughly ten times the tokens.
¶12 is a paragraph position, not a link. Quote it as plain text beside the source URL. A URL with #p12 glued on will not resolve.

Read the whole thing when you need it

Some work needs the article, not an answer: a translation, a rewrite, a close read. smry returns it as a window sized to a token budget, plus a marker for where it stopped. Hand the marker back and your agent continues from there.
read this whole thing and give me a section-by-section outline

What smry can open

Everything on that list answers a question the same way an article does. The format stops mattering once smry has the text. When a source is worth reading every time it publishes, follow it rather than pasting its links one by one.

A forty-page PDF you need one number from

Reading the PDF costs tens of thousands of tokens. Asking it a question costs hundreds.
in this filing, what did they report for free cash flow? https://harbor.example/annual-report.pdf
search_article · 1 call · returned 1 passage from a 40-page file

A conference talk you do not have an hour for

Paste the YouTube link and ask. smry reads the transcript, so a forty-minute talk answers a question in one call.
does the speaker in this talk say anything about hiring? https://www.youtube.com/watch?v=example
An .xlsx attachment, a CSV export, a slide deck. smry reads it from a URL, so give it the link the file lives at. If the file is only on your disk, paste the text into the prompt — every tool your agent has takes a URL or a string, never raw bytes.
here is the csv. which rows have a renewal date before March?
File bytes go in through the HTTP upload endpoint instead, which no agent tool calls. See Work with the library.

What cannot be read

smry does not get past a hard paywall, a sign-in wall, or a private network. That holds for sites you personally subscribe to. When a page cannot be read, smry returns an error naming the reason. It never returns partial text dressed up as the article, and never guesses at what the page said.

Questions

100 MB is the upload ceiling. Anything smry has to parse for text tops out at 25 MB, so a 40 MB PDF is rejected rather than half-read. Split it, or give smry the source URL instead of the file.
Every passage carries its paragraph anchor and its source URL, so you can check any claim against the page. The text itself was written by whoever published it, so treat it as source material rather than as instructions to your agent.
You get an error that says which kind of failure it was: a bad or unsupported URL, a publisher that timed out, or a budget you have spent. A retry helps on the second; it never helps on the first.
Any client you have connected — Claude Code, Codex, Cursor, Claude Desktop, or your own code. See Connect an agent.
A free account gets 60 operations a minute and 2,000 a day, of which 10 a minute and 100 a day can be fresh extractions. A page smry has already read comes from cache and costs no extraction. Paid plans raise the ceiling.

Next

Keep what matters

Save the pieces worth returning to, in a library only you can see.

Ask your library

Search everything you have kept, in one question.

Read a URL

The three read operations, their shapes, and their costs.
Last modified on September 13, 2026