See the clauses you were going to skim.
Drop a contract and get an outline, the dates and amounts in it, and the clauses people are told to look for — each one quoted from the page it was found on. The scan is free, needs no account, and reads the file in this tab. A written summary is optional and paid, and the page says what it sends before it sends it.
Up to 300 pages are read. The file is parsed in this tab and never sent anywhere, so there is nothing to delete afterwards.
- The text layer is read here
pdf.js turns each page into text in this tab. If a page is a scan with no text layer, it is named as unreadable rather than passed over silently.
- 27 rules look for clauses
One regular expression per clause type: renewal, liability, indemnity, arbitration, governing law, termination, assignment, notice periods and the rest. Fixed, so the same document always produces the same report. A short second list names the clauses an agreement usually has that none of the rules matched, so the gaps are visible too.
- Every line shows its working
Each finding quotes the sentence that matched, with the match marked, and the page it came from. The scan never paraphrases, so it cannot paraphrase wrongly. Anything written in prose is the separate summary, which is optional, paid, and labelled as a model's reading.
- It cannot read a negative
A rule that looks for “indemnif” matches “neither party shall indemnify the other” just as readily as an actual indemnity. The quote is always shown so this is visible, but it means a finding is a pointer to a sentence, not a conclusion. The list of clauses it did not find carries the same caveat: a clause worded differently is invisible to it.
- It is not legal advice
Nothing here judges whether a clause is fair, lawful or enforceable, and no output should be relied on as if it did. A lawyer reads meaning; this reads wording and shows you where to look.
- Scanned pages are invisible
There is no OCR. A contract that arrived as photographs of paper produces no findings, and the pages are listed as having no text layer so the empty report cannot be mistaken for a clean one.
- Headings are a guess, not a table of contents
The outline is recovered from numbering and capitalisation in the text. Embedded PDF bookmarks are not read, and an unusually formatted document may yield few headings or none.
- The written summary is a model, not a lawyer
The summary is generated text and can misread a clause that is worded unusually, so it is a starting point for reading, never a substitute. It is given only the first 30,000 characters of the text and the clause titles the scan found, it is told not to advise, and neither the text nor the summary is stored. Ticking the box is the only thing that sends anything.
Signing happens on this device too.
Up to four signers, colour-coded, a checklist that will not release the download until every field is signed, and a certificate written into the exported file.