An agent-first patent-prosecution assistant that runs on your machine, not the cloud.

An open-source desktop app where an AI agent drafts and redlines patent documents — office-action responses, claim amendments — directly in your own .docx files, as native Word tracked changes you accept or reject. It works inside a folder you already have, in open formats, readable without the app.

• Drafts in Word files, as tracked changes. On a Patrick-created .docx, the agent reads the document and proposes edits and comments as native tracked changes you accept or reject in the built-in editor. Your own files (PDFs, your .docx) are read-only — it offers to make an editable copy rather than touch the original.
• Brings your documents into the conversation. Pin a PDF or Word document and its text is sent to your chosen AI provider as context for the chat — a PDF as its native text layer, or on-device OCR for scans.
• Side-by-side workspace. View a source PDF or .docx, your editable draft, and the chat alongside one another.
• Primes the agent with verbatim law. Grounds answers in the exact wording of the EPC, EPO Guidelines, PCT-EPO Guidelines, and Boards of Appeal case law — pulled in by citation, or found by a search when you don't have the reference.
• Retrieves prior art. Pulls the full text of a patent (as Markdown) from Google Patents or the EPO's Open Patent Services; an optional per-chat web search rounds out research.
• Claim charts. Build a claim-limitations × prior-art table with a per-cell verdict, citations, and reasoning.
• In-document search. Hybrid semantic + keyword search over any open document.
• Your model. Bring your own key — Anthropic, OpenAI, Google, or via the Vercel AI Gateway — and pick the model per chat.

• Windows only, unsigned, with no auto-update yet (macOS and signing are planned).
• The agent edits Patrick-created drafts; it never modifies your original files — it proposes an editable copy instead.
• Headers and footers are view-only — shown faithfully, not yet editable.
• You can see the entire system prompt, but only the instruction layer you author (your profile prompt) is editable — the tool and agent scaffolding is fixed (visible, not editable).
• The law dataset is European (EP) focused.
• Documentation is sparse — in-app help and guides are still thin.
• Error handling is basic — not yet robust or comprehensive; failures can be abrupt rather than graceful.
• Not battle-tested — limited real-world mileage so far; expect rough corners.

• Open — Apache-2.0, and your work lives as plain .docx/.pdf in your own folders. No proprietary database, no lock-in; open it in Word tomorrow without Patrick.
• Transparent — you see exactly what the agent is told and doing: the full system prompt, its reasoning, every tool call, the documents in context, and the running cost. You author the instruction layer; the tool scaffolding is locked, but nothing is hidden. The agent proposes; you decide.
• Yours — everything stays on your machine, including your AI keys (bring your own — Anthropic, OpenAI, or Google). There is no Patrick server; your keys talk to your chosen provider, and nowhere else.

Private by design, not by policy — there's no server to trust with your clients' privileged work, because there isn't one.

Patrick is an unsigned Windows desktop app (alpha) — grab the installer from the Releases page. On first launch Windows SmartScreen may warn "Windows protected your PC"; click More info → Run anyway. You bring your own AI provider key and pay that provider directly for usage.

A pnpm monorepo. Each app/package, and what it's built with:

apps/
  frontend/   React 19 + Vite + Tailwind v4 + shadcn — the webview UI (desktop today; web/cloud later)
  api/        Hono on Bun — the local backend, compiled to a Tauri sidecar binary
  desktop/    Tauri — the desktop shell (frontend webview + api sidecar)
  site/       Next.js — the marketing + docs site
packages/
  shared/        TypeScript types, the model catalog, prompt tokens (frontend + api)
  ui/            @patrick/ui — the shared design system: shadcn primitives + stone/emerald tokens (frontend + site)
  law/           the EP law dataset + retrieval (EPC, EPO Guidelines, case law)
  benchmarking/  a grounding benchmark for the agent's legal accuracy

Across the repo: TanStack Router + Query, Vercel AI SDK v7 (Anthropic / OpenAI / Google / Gateway, bring-your-own-key), Biome, strict TypeScript, and bun:test.

pnpm install
pnpm dev              # frontend + API (browser)
pnpm dev:desktop      # Tauri desktop app
pnpm check            # typecheck + lint + dead-code check
pnpm test             # run the test suites (bun:test, from the repo root)

AI is bring-your-own-key — set your provider key in a profile inside the app. Nothing is sent anywhere except the provider you choose.

See CONTRIBUTING.md for the full workflow — branch → PR → merge, the changelog, dependency bumps, and releasing.

Apache-2.0.