Straight answers

Confidentiality

Where does my client's specification go?

When you use a model running on your own computer, your application documents, the text read from them, and your drafts are processed on that computer and are not sent to us or to an AI provider. If you choose a cloud provider, the text the model needs is sent to that provider under your account, and the app names the provider before anything is sent.

Apart from processing documents, a free build sends anonymous usage reports, which say how the app is used and carry no text from your work; DO_NOT_TRACK=1 turns them off. Every other connection is one you ask for or turn on in Settings: to check for updates, to run a patent search you request (the search describes the invention, and goes to SerpApi or your own Google Cloud project), to fetch a cited reference by its number, or to get a file history from the USPTO. Its log file stays on your computer. Every connection it makes.

Running locally limits where your documents go. Whether a tool is appropriate for a particular matter, including a privileged one, remains your professional judgment.

So which should I use?

A model on your computer, when the documents must not leave it. A cloud provider is faster and its strongest models draft better. Whether a particular provider suits a particular matter is your call, and it turns on the client, the engagement terms and the provider's own commitments.

What does the USPTO's 2024 notice on AI tools ask, and what does running locally cover?

The notice finds no prohibition on drafting with AI tools and no general duty to disclose their use; it applies the existing rules. On confidentiality, it says that entering an invention into an AI system can disclose it, that such systems may retain what is entered, and that servers outside the United States raise export questions. With a model on your computer, the specification is not entered into anyone else's system; with your own provider key, those points apply to that provider as to any other AI tool.

Its central point applies either way: whoever signs a paper filed with the USPTO must have reviewed and verified it, and relying on an AI tool's accuracy is not the reasonable inquiry the rules require. The passages each candidate cites and the critique's findings are there for that review; they do not make the inquiry. The notice, quoted from its source, is on patent-actions.com.

Does the app phone home?

Not for drafting, ever. A free build does send us usage reporting: which steps ran, whether they finished, which kind of model ran them and the hour, under a random install number, and never your work or your clients'. DO_NOT_TRACK=1 turns it off at no cost and deletes what is waiting (what a report holds). Apart from drafting, those reports, and the reference fetches and searches you ask for or turn on, these are the calls it makes:

• A check for a newer version, when you press Check for updates or run continuation-drafter update, or each time the app opens if you turn that on in Settings (it is off unless you do).
• With your own USPTO Open Data Portal key, a published application's file history, when you press Get the file history (sends only the number), and a check for new papers in each matter each time the app opens, if you turn that on in Settings (it is off unless you do). Each request carries an application number and your key, and nothing else.
• A license renewal, which only a paid license makes. No license is on sale, so today that call is never made, and CD_NO_LICENSE_RENEWAL=1 switches it off regardless.

The update check is one request to GitHub's public releases page carrying nothing about you, and offline mode refuses it. An update is installed only after its signature and checksum verify, so a tampered release is refused rather than installed.

For a firm's IT or security review, every connection the program makes, where it keeps its files and how a download is signed are on one printable page.

Is there any analytics on this website?

Yes: Cloudflare's web analytics, which records page views, page timings, and coarse data such as country and browser. It is cookieless, it does not fingerprint your device, and it builds no profile across visits or across sites. A cookieless one was chosen deliberately, because a practitioner's browsing is itself suggestive of client work.

The website and the app are measured separately. No specification and no draft ever reaches this site: it renders pages and holds nothing else. The program you download sends the usage reports and makes the calls described above, and no others.

What it does, and where it stops

What is it good at?

Drafting a broad set of continuation claim candidates from a specification you have already filed, quickly, so you start by editing rather than from a blank page. It also flags potential issues in each claim (antecedent basis, support, indefiniteness) with suggested revisions for your review, which in our own testing was the most useful thing it produces. The numeric score is the weakest part, and the benchmark page says so.

If what you want is the doctrine rather than the app, that is a different thing and it is also free: the continuation course sequences those same authorities, a question answered before anything else, and then the steps in the order a family is built. It sells nothing and requires no account.

Can it read the prosecution history?

Yes. Attach the parent's Patent Center "download all documents" PDF. Patent Center's pages are scanned images, so the app reads them with an OCR engine on your own machine (tesseract, if installed, or a PDF you have already run text recognition on), and no page leaves the machine for that step. A long history downloaded in parts is read as one, and any single Patent Center download is accepted. It reports what each paper says: the rejections, the references each relies on, the examiner's reasons for allowance, and whether each limitation of the allowed claims was filed with the application or added during prosecution. With the file history attached, each candidate embodiment carries a sentence relating it to that record: the limitations of the claims as allowed it builds on, each placed in the claims as filed or a round of prosecution, and the examiner's reasons for allowance, every claim, round and quotation checked against the record. It does not assess patentability, and it does not say why a limitation was added, whether an amendment was necessary or whether a rejection was correct; that is the practitioner's reading. How well it reads an Office Action was measured on real files. The same program also helps draft the reply to an Office Action, which patent-actions.com describes.

It does not assess patentability, novelty or non-obviousness

It is built not to. It does not assess patentability, novelty, non-obviousness or prior art. Those are legal determinations. This app drafts claim language and scores drafting quality against the specification you supply: how well it supports a claim's elements and their particular combination, how clearly the claim is worded, and how it differs from the claims already filed. The scoring prompt refuses patentability questions on purpose, and the refusal is printed in the output rather than hidden.

The authorities those determinations actually rest on are published separately and free of the app: the decisions, statutes and rules, each quoted from its primary source. Reading them is how you make the judgment yourself; nothing there assesses anyone's claims either.

The output is not filing-ready

It is Beta software, and every output says it is not filing-ready. It is drafting input for a registered practitioner who reviews every limitation against the specification and exercises their own judgment. Treat it as a first pass from a capable but unsupervised junior: worth having, never sent out unread.

It does not replace a registered practitioner

It is a drafting aid built for registered patent attorneys and agents, and for people working under their supervision. It does not practice law, and nothing it produces is legal advice.

I am an inventor drafting a continuation of my own application. Can I use it?

Yes. Anyone can download and use it, and an applicant may prosecute their own application (37 CFR 1.31(a)). It is built for registered patent attorneys and agents, so it uses their terms and assumes their review: whoever signs a paper filed with the USPTO certifies that it rests on an inquiry reasonable under the circumstances (37 CFR 11.18(b)), and the USPTO has said that relying on an AI tool's accuracy is not that inquiry. An applicant that is a company, or that includes a party domiciled outside the United States, must be represented by a registered practitioner (37 CFR 1.31(a)).

The USPTO offers inventors free help: its Pro Se Assistance Center, the Patent Pro Bono Program, law school clinics, and a directory of registered practitioners.

Does it search for prior art, or compare my claims against the references?

It searches only when you ask, and it does not judge what it finds. It can fetch a reference the examiner cited by its number, from Google Patents or, with your own USPTO Open Data Portal key, from the Portal, which also reads the copy of a cited foreign reference in the application's file history; and it can run a patent search you ask for in your own words, on your own Google Cloud project or SerpApi key. Each asks before it runs unless you turn that off in Settings, and offline mode refuses both. A search lists the documents it returned and the paragraphs where its terms appear, and says nothing about whether anything is new. A fetched reference is shown at the place the examiner cited, beside the citation, with the paragraphs just before and after it that continue its discussion, and the description's discussion of a cited figure, marked as context. Beside a cited figure, a button such as Open FIG. 5 opens the reference's printed copy at the drawing sheet that shows it. A reference it cannot get says why, and you can upload it where it is cited. Whether a document is prior art to your claims, or describes a limitation, is the practitioner's reading: the app does not assess prior art.

Running it

Do I need to use the Terminal?

On a Mac, no. Download the app from the download page, open it, drag Patent Practitioner Tools to your Applications folder, and double-click it. The drafting interface opens in your browser. macOS asks once whether you are sure you want to open something downloaded from the internet; the button is Open.

On Windows and Linux there is no app bundle, so you start it once from a terminal, or on Windows by double-clicking the .exe, which opens a console window. One command, ./continuation-drafter serve, and the same interface opens in your browser.

Everything the command line does is in that interface. Command-line binaries are published for macOS, Linux and Windows if you would rather work at a prompt, and they are the right choice for a script or CI.

The app has no Dock icon, because it runs no window of its own: the interface is the browser tab it opens. Press Stop in that tab to shut it down.

What does it cost?

The app is free to download and use. If you run local models, there is no marginal cost at all. If you use a remote provider, you pay that provider directly with your own key; nothing routes through us and we do not mark it up.

Either way, every run ends by reporting what it used and what it cost. On your own machine that report sits beside what the same work would cost on a leading hosted model at its list price; through a provider it is the cost that provider reported. The web interface also keeps a running total for each matter.

What setup does it need?

To open it, none, and installing it takes under a minute: on a Mac, drag the app to your Applications folder. To read scanned pages, an OCR engine (tesseract) on the same computer. To draft, a model, either on your computer or through your own key at a provider.

On your computer, the model runs through Ollama or another server that speaks the OpenAI API, such as LM Studio or llama.cpp; the setup page names the one that fits your machine. A local model is a download of several gigabytes, once, which is under an hour on a fast connection. Through a provider, your own key at OpenRouter, OpenAI or Anthropic needs no download. Choosing a model or pasting a key takes a few minutes.

Try the sample opens with neither: a reply's sample is already read, so a model is needed only once you ask for a draft, while continuation claims need a local or cloud model to get started, because their first step is finding embodiments.

Can I try it without client documents?

Yes. Try the sample opens a sample matter for either task: the public Office Action of application 18/541,216 for a reply, or a short published-style application for continuation claims.

The public papers of 18/541,216 (its Patent Center file history, specification and allowed claims) can also be downloaded and added yourself, which is the walkthrough on the download page.

What hardware do I need?

For local models, memory is the binding constraint, and it depends on your specification's length rather than its complexity. The setup page asks what machine you have and names a model that fits, with the measured tradeoff.

How long does a draft take?

Across 131 measured runs over 10 real patent families and 15 remote models, the median was 78 seconds to a full set of about 13 claims, and 84% of runs finished inside three minutes. Those runs used remote models; on your own machine it is slower, by an amount that depends on your hardware.

Those times are for a whole set drafted in one run, which is what Draft straight through does in the app. Continuation claims in stages spend the time where you decide: Discover embodiments drafts nothing, the first claims toward the one you choose are three, and the full set is drafted only once you have shaped them.

My specification is enormous. Is that a problem?

Less than you would expect. Across those same runs, specification length did not predict drafting time (the correlation was -0.118 over specifications from about 17 pages to about 240). The longest specification tested drafted faster than a third of the smaller ones. Which model you choose matters far more than how long your document is, though the model's context window does have to fit the specification, which is what the setup page is for.

Does it work with Word?

Yes, both ways. Every paper can be added as a Word file, as well as a PDF or text, and struck-through text in one is read as deleted. Claims save as Text, Word or PDF wherever they are shown, and a reply downloads as Word, from Reply in Word or Report of the action in Word; notes on claims or a reply are never written into a download.

Claim numbering is kept exactly as drafted and never renumbered, since a continuation's claims are commonly numbered on from the parent. Word is the format the USPTO expects for the description, claims and abstract; filing them in another format carries a surcharge. From the command line, continuation-drafter export --draft claims.txt --out claims.docx, or pipe a drafting run straight into it; LaTeX works too, with --out claims.tex.

The project

Is this open source?

No. The binary is free to download and use, and it is proprietary. The source is not published. Each release publishes THIRD_PARTY_NOTICES.md beside its release notes, with the license of every third-party component it is built with.

Why is it free?

It is free because it is in Beta, and the point now is for practitioners to use it and say what is wrong with it. A free build reports anonymous usage, never your work or your clients', so we can see which steps are worth keeping and where people get stuck, and DO_NOT_TRACK=1 turns that off at no cost. A paid tier is possible in the future, but there are no plans for one at this time.

Who is behind it?

Patent Practitioner Tools is built and owned by Obviously Not, Inc.: about the company. It came out of internal R&D on continuation drafting and was extracted into a tool once it drafted end to end. The benchmark that the benchmark page reports is the same one used to decide which models ship in the guided setup.

Why should I believe the benchmark?

Because you can check it. The method is published, the models and families are named, and the limits are stated on the page rather than in a footnote: a single AI judge cannot reliably rank near-equal drafters, so the scores separate gross differences and nothing finer. That is a narrow claim, and one you can verify.