Frequently asked questions

Patendra is a local-first patent research and drafting workspace. A matter connects technical disclosure, configured prior-art sources, claim refinement, examination issue review, score inputs, and an editable first draft. It ships as desktop software for macOS and Windows and can also run from source.

Patendra is designed for patent practitioners, in-house IP teams, inventors, and R&D groups that want a more structured first pass. Every output requires qualified professional review before filing.

A matter moves through disclosure structuring, configured prior-art retrieval, claim refinement, examination issue review under §102, §103, and §112, scoring, and editable first-draft assembly. Intermediate artifacts stay in the local matter record so important conclusions can be reviewed with their inputs. See how it works.

The first run downloads a sentence-transformer embedding model used by the local vector store, so it needs an internet connection once and takes a little longer. Later runs reuse the downloaded model.

Anthropic Claude and Google Gemini. You bring your own API key for either provider (or both) and can switch models per run. The model picker disables providers that have no key configured, instead of letting a run fail.

A chat model drafting freehand has no retrieval gate: it can invent prior-art citations, assert novelty it never checked, and produce claims with structural defects an examiner would flag immediately. Patendra separates reasoning from evidence: rejections may only cite retrieved, ID-verified references, claims pass deterministic structural validation (antecedent basis, dependencies) before any model refines them, and the score is computed from measurable signals rather than stated by the model. See the full feature breakdown.

It complements one rather than replacing it. Patendra gives you a fast, repeatable, evidence-cited search and simulated examination for pre-filing triage. A professional searcher or attorney adds human judgment, deep non-patent-literature coverage, and a legal opinion you can rely on. Many users run Patendra first and bring counsel a much better-informed question.

In a local data directory on your machine: ~/Library/Application Support/Patendra on macOS, %APPDATA%\Patendra on Windows, or wherever PATENDRA_HOME points. That directory holds your API keys, runs, learning data, and vector store. Nothing is uploaded to Patendra. There is no Patendra cloud. The complete outbound-call map is on the security page.

Patendra does no training on your data. The cross-run learning loop is local statistical calibration stored in your own data directory, and it never leaves your machine. LLM calls do send disclosure text to your chosen provider (Anthropic or Google) under that provider's API terms, so review their data-use and retention policies before running sensitive material.

Partially. The first run needs an internet connection to download the embedding model. After that, heavy features degrade gracefully offline: the local vector store keeps working, while LLM calls and remote patent-data searches naturally require connectivity.

Yes. The server binds 127.0.0.1 by default and refuses non-loopback addresses without PATENDRA_API_TOKEN. It fails closed. With a token set, every API endpoint requires it, and the dashboard uses an HttpOnly session cookie so browsers never hold the raw token. A Docker image is available. Deployment recipes are in the docs; the hardening list is on the security page.