Answering an Office Action, end to end
Three open-source tools for patent prosecution. Self-hosted — you run them on your own machine with your own Anthropic API key. There's no hosted instance and no account to create.
Your documents go from your machine straight to the Claude API under your own key — there's no server in between holding them.
The OA Agent in 40 seconds
Three tools, three jobs
Responding to an Office Action means reading the art the examiner cited, working out which rejections to fight and which to amend around, and making sure the amendments hold up. Each tool covers one of those. They run independently — there's no shared pipeline between them.
Read the cited art
Patent ReaderEnter a publication number and get claim dependency trees, reference numerals cross-linked to the figures, bounding boxes on the drawings, and multi-term search. Built for reading a reference the examiner cited against you.
Decide argue or amend
OA AgentAn AI agent that follows the examiner's paragraph and column citations into the actual prior art, maps each claim limitation to the evidence relied on, and recommends argue or amend. Every conclusion cites a passage you can open and check.
Check your amendments
Antecedent BasisPaste a claim set and get the §112 problems that creep in when claims are edited under time pressure — missing antecedent basis, ambiguous pronouns, unclear "such" phrases, prepositional scope, and silent reintroduction.
How it works
Built to show its work
The point is that you can check everything it tells you. Nothing it concludes is something you have to take on faith.
It opens the documents
When the examiner cites column 5 of a reference, the agent opens that reference and reads column 5. It works from the files in front of it, not from what it remembers about the patent.
You watch it work
You see which document it has open and which passage it's reading while it happens, rather than a summary afterwards. If it goes down the wrong path you'll know immediately.
Every claim is traceable
Each conclusion points at the passage it came from. Click it, read the source, and decide for yourself whether the reasoning holds.
Not everything is AI
The antecedent basis checker reads claim grammar directly rather than asking a model. It gives the same answer every time and can't invent a finding that isn't there.
Run it yourself
Clone the repo, follow the setup guide, and it runs on your laptop.
Setup guide on GitHub