Draft your Office Action response with AI, on your own computer
Patent Practitioner Tools lists each rejection, objection, and requirement in the action with the examiner's citations. You choose how to answer each one, and it drafts suggested language for that approach, which you adopt, edit, or discard before the reply goes to Word. You start by editing a draft rather than from a blank page.
Free to download and use, with no account and nothing to activate. For macOS, Windows, and Linux, with a model on your computer or your own provider key.
AI-assisted drafting: you choose the strategy, verify the facts and authorities, and remain responsible for the submission.
Address the entire Office Action, one item at a time
Work through the action's rejections, objections, and requirements, each in its own workspace. Review and revise the suggested text, then add the parts you choose to your reply. Your notes stay out of the exported reply.

- Next steps holds back a statement that a reference does not describe something until you confirm what you reviewed.
- The items, filterable by state, from no approach chosen to reviewed.
- Each item's row: its statute and claims, the references by name, its approach and its counts of units.
- Response approach is yours to select for each item; the app generates draft text based on your selection.
- From the record is written from the prosecution record without an AI model; Suggested text is AI-generated; Review marks text a check flagged. All of it requires your verification.
- An amendment added to the reply writes its words into the claim it names, in your proposed claims.
- Not added marks a unit waiting on Add to draft; nothing drafted is in your reply until you add it, and Edit and Dismiss keep its history.
The action and the reply, pictured, and each step on the public sample with what to try first
What it does for the reply
The record, as you have it
Add the Office Action and the papers around it as you have them: one by one, as a Patent Center download, or, with your own USPTO Open Data Portal key, as a published application's file history the app gets for you. Each paper is sorted by what it is and read on your computer.
Every item, in the examiner's words
Each rejection, objection, and requirement the action states, with its claims, ground, and references, marked new, maintained, modified, or withdrawn against the action before. The action's dates are estimates, with how they were counted: verify them against the Office Action and your docket, which remains the system of record.
The claims, against the citations
The whole claim set, each independent claim with the examiner's citations against its limitations and its dependent claims beneath it. An amendment you add is written into your proposed claims, marked up as the claim listing marks it, and the drafter argues the claims as amended.
A draft built one part at a time
You select the response approach for each item; the app generates draft text based on your selection, in drafted paragraphs it calls units. Text marked From the record is written from the prosecution record without an AI model; Suggested text is AI-generated; Review marks text a check flagged or whose review scope you have not confirmed. All of it requires your verification. Each unit goes into your reply with Add to draft, or is edited or dismissed.
Every quotation checked against the paper
Each quotation the app shows or drafts, from the action, the specification, or a reference, is looked for in that paper. One it cannot find is marked, and a drafted paragraph that quotes it is withheld, so a draft cannot quote words the paper does not contain. Each cited reference shows its dates beside the earliest date the application claims; whether it is prior art is yours to decide. What the USPTO expects of a signer who uses AI.
The reply, in Word
Read the reply as one document, with checks over what you wrote. Reply in Word is laid out as a filed response, in a sample layout or in the form of a reply you upload, with the claim listing marked up, and notes stay out of the export.
Around the reply
A report of the action in Word, each item in the examiner's words with the claims it affects and the period the action states, for you to send to a client or adapt. An interview agenda to paste into the Applicant Initiated Interview Request. The references of record for an IDS, with the paragraph of 37 CFR 1.97 an IDS filed on a date you enter falls under. Excess-claims fees for the proposed claims. After a final action, the paths the rules provide, each with its fee. Each states what the record says, never what to do.
Try it on a public Office Action first
Open the app and press Try the sample under Draft a response to an Office Action. It opens the non-final rejection of application 18/541,216, public record, already read, so you can see its items and choose approaches before any model is set up, and draft them once one is. No client material is involved.
Running it on your own computer
What does it cost?
It is free to download and use, with no account and nothing to activate. With a local model there is no cost per use; with a provider key you pay that provider directly, with no markup from us. More on cost.
What setup does it need?
To open it, none. To read scanned pages, an OCR engine (tesseract) on the same computer. To read what an Office Action states, and to draft, a model: one on your computer, through Ollama or another server that speaks the OpenAI API, such as LM Studio or llama.cpp; or your own key at OpenRouter, OpenAI or Anthropic. Try the sample opens with no model, because its Office Action is already read; a model is needed once you ask for a draft. More on setup.
What stays on my computer?
With 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. A model your Ollama server would send on to another machine, such as an Ollama cloud model, is refused. Other network activity, such as an update check you ask for, is listed on /security. More on where the papers go.
Measured, not promised
What it was measured on, the method and the runs that did worse are published, so you can check them.
qwen3.5:9b on a laptop
0 references named that the paper does not contain, across every run
Drafting replies to 10 real Office Actions, gemma4:31b, running
locally, raised 15 of the 87 points the filed replies made, beside
14 for gpt-5.6-sol and 20 for
claude-opus-5.5.
In 77 scanned Office Actions, the app attached an examiner's citation to 79% of 1,287 claim limitations, with no AI model. Whether it was the right one was checked on 20 of those actions, on the 190 limitations where two AI models from different vendors quoted the same citation: 86% of the citations it attached there matched theirs, and 77% of those limitations got the right one. That reference set was made by models, not checked by a practitioner.
Reading real file histories: the table, the method and its limits · The drafts against the filed replies
The rules an Office Action runs on, quoted
Guides to the questions an Office Action raises, built on the authorities themselves: the rules, manual sections, notices and decisions, every quotation checked against its source.
Start with the guides, or go straight to the authorities. What each authority says, not advice about any matter.
Questions before you install
- What reaches a provider if I use my own key?
- Can I try it without client documents?
- Does it work with Word?
- Does it say whether a rejection was right? No: the reply is yours to write, sign and answer for.
- Who is behind it?
Try it on an Office Action you have answered
Add one you have already answered, ask for a draft of the reply, and read each drafted part against what you filed.
When the next step is a continuation
The same app drafts continuation claims from the specification, using the file history it has just read. Patent Practitioner Tools is made by Obviously Not, Inc.
Allowed claims, the limitations added to reach them and the examiner's stated reasons are the record a continuation is drafted against. The drafting side, its benchmark across 51 models and its free course are described at patentcontinuation.com. To choose a model for your machine, start with set up or what was measured.