Claude Patent Creator

MIT License Python 3.9+ MCP Server PyTorch Status

An AI-powered patent creation and analysis system for Claude Code.

I built this because I needed to file a patent myself. I used AI to build the system, used the system to file the patent, and it worked. Now it's open source so anyone can use it. You do not need to be a lawyer or a programmer. If you can describe your idea in a chat window, you can use this.

In plain terms, this tool lets you:

• Ask patent-rule questions in plain English. "Do my claims have to be numbered?" "How long can the abstract be?" You get the actual rule, with a citation, in about a second. (Under the hood it searches the MPEP, the examiner's rulebook, plus the patent statutes and regulations.)
• Find out whether your idea is actually new. Search more than 100 million existing patents worldwide for anything close to your invention. This is called a prior art search, and it is the single most important step — it is how you learn whether a patent is even worth pursuing.
• Get your draft checked before a human ever bills you an hour. The tool reads your claims and application the way an examiner would and flags the mechanical problems: unclear wording, missing sections, an abstract that is too long. It also tells you which checks it did not run, so a clean report means what it says.
• Make the drawings. Patent-style diagrams and flowcharts from a description, no design software.
• Go from "here is my idea" to a reviewable draft application. A guided workflow takes your description (or even your source code), hunts for prior art, drafts claims and a specification, checks everything, and hands you a package a patent professional can review quickly — for US or European filing.

New to the vocabulary? Every term this page uses is explained in the Glossary in plain English.

What this tool will and will not do — read this first

This tool is built for people who cannot afford a mistake, so it is honest to a fault. Know what you are getting:

• It will tell you "no." If the search finds your idea already exists, the tool's job is to show you the evidence and save you the filing fee — not to flatter you into spending money. A well-reasoned "this is not new, and here is why" is the tool working, not failing.
• It never says "verified" loosely. Every automated check reports what it ran and what it skipped. If a report looks clean, you can trust that the words mean exactly what they say.
• It does not file anything for you. You review, you decide, you file. The tool prepares; you commit.
• It is not a lawyer and does not give legal advice. What it produces is a well-prepared draft plus organized evidence — built so that a patent professional's review is fast and cheap instead of slow and expensive. For anything you intend to file, have a registered patent practitioner look at it. The point of this tool is to make that look-over affordable, not to skip it.
• Some judgments are human-only. Whether your invention is "non-obvious" and whether it is the kind of thing patent law protects at all are legal judgments no software makes reliably. The tool arms you for those conversations; it does not have them for you.

Table of Contents

• What this tool will and will not do
• Quick Start
• What Can I Actually Do With This?
• How It Works
• Installation Options
• CLI Commands
• MCP Tools Reference
• Skills, Agents, and Slash Commands
• Configuration
• Requirements
• Architecture
• Performance
• Known Issues
• Glossary
• Roadmap
• Contributing
• Credits
• License

Quick Start

Never used Claude Code? It's Anthropic's AI assistant that runs in a terminal or as a desktop app — install it first (a paid Claude subscription is the only cost to start). Then come back here; setup is one command and the tool talks you through the rest.

Pick the path that fits your setup. All three get you to the same place.

Option A: Claude Code Plugin (Easiest)

If you're already using Claude Code, this is the fastest way in:

# Add the marketplace and install
/plugin marketplace add RobThePCGuy/Claude-Patent-Creator
/plugin install claude-patent-creator-standalone@claude-patent-creator

# Run setup
/claude-patent-creator-standalone:setup-patent-system

Option B: One-Line Install

pip install git+https://github.com/RobThePCGuy/Claude-Patent-Creator.git && patent-creator setup

This handles everything automatically: installs dependencies, detects your GPU, downloads MPEP PDFs (~500MB), builds the search index, and registers the MCP server with Claude Code. Restart Claude Code when it finishes.

Option C: Manual Install

git clone https://github.com/RobThePCGuy/Claude-Patent-Creator.git
cd Claude-Patent-Creator

# Optional: use a virtual environment
python -m venv venv
source venv/bin/activate  # Linux/macOS
venv\Scripts\activate     # Windows

pip install -e .
patent-creator setup

Verify It Worked

After any install path, run:

patent-creator health

You should see a status report showing which components are ready. If something's off, the output will tell you what to fix.

What Can I Actually Do With This?

Here are some real examples. You can type these directly in Claude Code and the right skill or tool kicks in automatically. Say it in your own words — the "what to say" column is a starting point, not a required incantation.

How It Works

Everything from here down gets progressively more technical. You do not need any of it to use the tool — the sections above plus the Glossary are enough. This part is for developers, patent professionals, and the curious.

The system has two modes that can work independently or together:

MCP Server is the engine. It exposes 20+ tools that any MCP-compatible client (Claude Code, Claude Desktop, etc.) can call programmatically. These tools handle search, analysis, and diagram generation.

Claude Code Plugin adds the interactive layer. Skills activate automatically based on what you're doing. Agents handle long-running tasks in the background. Slash commands give you quick access to common workflows.

Under the hood, patent regulation search uses a hybrid approach: FAISS vector search finds semantically similar content, BM25 lexical search catches exact terminology matches, and a cross-encoder reranker sorts the combined results by relevance. Patent search goes through Google BigQuery's public patent dataset.

You (Claude Code) ──> MCP Server ──> Search / Analysis / Diagrams
                           │
            ┌──────────────┼──────────────┐
            v              v              v
     MPEP/USC/CFR     BigQuery        Graphviz
     (hybrid RAG)    (100M+ patents)   (diagrams)

Installation Options

python -m venv venv
source venv/bin/activate  # Linux/macOS
venv\Scripts\activate     # Windows

pip install git+https://github.com/RobThePCGuy/Claude-Patent-Creator.git && patent-creator setup

claude --plugin-dir ./Claude-Patent-Creator

CLI Commands

patent-creator setup             # Full setup wizard (downloads, builds index, registers MCP)
patent-creator health            # System health check (shows what's working and what isn't)
patent-creator status            # Same as health
patent-creator verify-config     # Check Claude Code MCP configuration
patent-creator serve             # Run the MCP server manually
patent-creator rebuild-index     # Rebuild the MPEP search index
patent-creator download-mpep     # Download MPEP PDFs only
patent-creator download-all      # Download all sources (MPEP + 35 USC + 37 CFR)
patent-creator check-bigquery    # Test BigQuery connection

MCP Tools Reference

Search

Getting the campaign workflow (skills)

The MCP server above gives Claude the TOOLS. The campaign workflow — the skill that runs mining, prior art, worth-it economics, drafting, and the red teams end to end — ships as a Claude Code plugin. Two commands:

claude plugin marketplace add RobThePCGuy/Claude-Patent-Creator
claude plugin install claude-patent-creator-standalone@claude-patent-creator

Re-run the install command after upgrades to refresh the skills.

Analysis

Generation

System

Skills, Agents, and Slash Commands

Skills (activate automatically)

You don't need to call these directly. Just describe what you want to do and the right skill kicks in.

Agents (long-running, work independently)

These run in the background while you keep working on other things.

To use them: "Create a patent for [your invention], use the patent-creator agent"

Slash Commands

/create-patent          # Complete patent creation workflow (55-80 min)
/search-prior-art       # Prior art search with novelty analysis
/full-review            # Parallel review (claims + spec + formalities)
/review-claims          # Claims-only 35 USC 112(b) analysis
/review-specification   # Specification-only 35 USC 112(a) analysis
/review-formalities     # MPEP 608 formalities check

Configuration

Environment Variables

Create a .env file in the project root (see .env.example):

# Required for BigQuery patent search
GOOGLE_CLOUD_PROJECT=your-project-id

# Optional: EPO OPS API (for EP patent full-text search)
# Free registration at https://developers.epo.org
EPO_OPS_KEY=your-consumer-key
EPO_OPS_SECRET=your-consumer-secret

# Optional API keys (for HyDE query expansion)
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...

# Optional settings
PATENT_LOG_LEVEL=INFO
PATENT_LOG_FORMAT=human
PATENT_ENABLE_METRICS=true
PATENT_MPEP_USE_HYDE=false
PATENT_MPEP_DEVICE=gpu
PATENT_OPERATION_TIMEOUT=300

# 1. Install Google Cloud SDK
# https://cloud.google.com/sdk/docs/install

# 2. Authenticate
gcloud auth application-default login

# 3. Set quota project (replace with your project ID)
gcloud auth application-default set-quota-project your-project-id

# 4. Set the environment variable
export GOOGLE_CLOUD_PROJECT="your-project-id"   # Linux/macOS
$env:GOOGLE_CLOUD_PROJECT="your-project-id"     # Windows PowerShell

# 5. Test it
patent-creator check-bigquery

# In your .env file
CLAUDE_CODE_GIT_BASH_PATH=C:\Program Files\Git\bin\bash.exe

# Correct
claude mcp add ... -- "C:/Users/YourName/venv/Scripts/python.exe"

# Wrong (will fail)
claude mcp add ... -- "C:\Users\YourName\venv\Scripts\python.exe"

Requirements

Minimum

• Python: 3.9 - 3.13 (3.14 is experimental)
• RAM: 8GB
• Disk: ~2GB (MPEP PDFs + search index)

Optional (but recommended)

• GPU: NVIDIA with CUDA 12.8 (makes indexing 5-10x faster) or Apple Silicon (2-3x faster)
• Google Cloud: Project with BigQuery enabled (for patent search)
• Graphviz: System package (for diagram generation)

Architecture

claude-patent-creator/
├── .claude-plugin/          # Plugin manifest and marketplace config
├── mcp_server/              # Core MCP server
│   ├── server.py            # FastMCP entry point
│   ├── mpep_search.py       # Hybrid RAG search engine
│   ├── bigquery_search.py   # BigQuery patent search
│   ├── claims_analyzer.py   # 35 USC 112(b) analyzer
│   ├── specification_analyzer.py  # 112(a) analyzer
│   ├── formalities_checker.py     # MPEP 608 checker
│   ├── diagram_generator.py       # Graphviz diagrams
│   ├── tools/               # MCP tool definitions
│   └── index/               # FAISS + BM25 index (git-ignored)
├── skills/                  # Claude Code skills (13)
├── agents/                  # Autonomous agents (10)
├── commands/                # Slash commands (11)
├── hooks/                   # Event-driven automation
├── scripts/                 # Testing and utilities
├── docs/                    # Additional documentation
├── pdfs/                    # Downloaded MPEP PDFs (git-ignored)
└── CLAUDE.md                # Full project documentation

For the complete architecture documentation, development workflows, and troubleshooting guides, see CLAUDE.md.

Performance

Resource usage: the loaded search index takes about 2-4GB of RAM and the index files are 500MB-1GB on disk. If you have a GPU, it'll use 1-2GB of VRAM for acceleration.

Known Issues

This project is a work in progress. Most features work, but expect some rough edges. Contributions, issues, and PRs are welcome.

Things to be aware of:

• PyTorch install order matters. Install PyTorch before sentence-transformers, or you'll end up with CPU-only PyTorch even on a GPU system. The setup wizard handles this, but it can bite you on manual installs.
• BigQuery requires a Google Cloud project with billing enabled. The patent data itself is free to query within the BigQuery free tier.
• Some diagram types need Graphviz installed as a system package (not just the Python bindings).
• HyDE query expansion requires API keys (Anthropic or OpenAI). It's optional and off by default.
• Windows users need Git Bash for the claude mcp add command. See Windows setup notes.

See CLAUDE.md for the full troubleshooting guide.

Glossary

If you're coming from the development side and patent terminology is new (or vice versa), here's a quick reference:

Roadmap

• Support for international patent offices (EPO, WIPO/PCT)
• Web interface for non-Claude Code users
• Claim dependency graph visualization
• Automated obviousness analysis (35 USC 103)
• Patent portfolio analysis tools
• Integration with patent drafting software

Contributing

This project is open to contributions. Since it's a work in progress, expect breaking changes and incomplete documentation. Issues and PRs are welcome.

See CONTRIBUTING.md for the development setup, branch naming, commit conventions, and code style guide.

Credits

Open Source Dependencies

This project builds on excellent open source work: FastMCP, FAISS (Meta AI Research), Sentence Transformers (UKP Lab), HuggingFace Transformers, PyTorch, rank-bm25, PyMuPDF, Graphviz, Pydantic, and Google Cloud BigQuery.

Data Sources

MPEP, 35 USC, and 37 CFR are published by the USPTO. Patent data comes from Google BigQuery's patents-public-data dataset (100M+ patents). Embedding models are BGE-base-en-v1.5 (BAAI) and MS-MARCO MiniLM-L-6-v2 (Microsoft).

Trademark Notice

"USPTO" is a registered trademark of the United States Patent and Trademark Office. This project isn't affiliated with, endorsed by, or sponsored by the USPTO.

Project Status

This project is in beta. I'm actively working on it, but not everything is polished and some features may not work as described. If you run into issues, open one on GitHub and I'll take a look.

For detailed documentation: CLAUDE.md | For security issues: SECURITY.md

License

MIT License. See LICENSE for details.

Built with Claude Code. The code is the output, but the real work is deciding what needs to exist and how the pieces fit together.