Claude Patent Creator
USPTO patent creation and analysis system with dual-mode architecture: MCP server for programmatic access + Claude Code plugin with skills, autonomous agents, and slash commands. Features hybrid RAG search (FAISS+BM25+reranking) over MPEP/USC/CFR, BigQuery access to 76M+ patents, automated compliance checking, prior art search, and diagram generation.
Project Status
This project is a work in progress and is not fully functional.
Contributions, issues, and pull requests are welcome. Feel free to explore, experiment, or build upon the code. No guarantees of stability or completeness are provided.
Features
Dual-Mode Architecture
| Mode | Use Case | Components | Access Method |
|---|---|---|---|
| MCP Server | Programmatic API access | 20 MCP tools | Any MCP client (Claude Code, Claude Desktop, etc.) |
| Claude Code Plugin | Interactive workflows | 15 skills + 10 agents + slash commands + hooks | Claude Code IDE |
Core Capabilities
Search & Retrieval (20 MCP Tools)
- MPEP/USC/CFR Search - Hybrid RAG (FAISS vector + BM25 lexical + cross-encoder reranking) across 500MB of USPTO regulations
- Patent Search - BigQuery access to 76M+ worldwide patents with full-text search and CPC classification
- USPTO API Integration - Real-time patent data retrieval and recent filings
- Prior Art Discovery - Automated novelty and freedom-to-operate analysis
Automated Analysis
- Claims Review - 35 USC 112(b) compliance: definiteness, antecedent basis, indefinite terms, claim structure
- Specification Review - 35 USC 112(a) adequacy: written description, enablement, best mode
- Formalities Check - MPEP 608 compliance: abstract length, title format, drawings, required sections
Content Generation
- Diagram Generator - Patent-style technical diagrams using Graphviz (block diagrams, flowcharts, system architectures)
- Patent Creation - Guided workflow for drafting complete USPTO-ready applications
Claude Code Plugin Features
- 15 Skills - Specialized expertise modules (setup, development, index management, troubleshooting, testing, patent review, search, diagrams, prior art)
- 10 Autonomous Agents - Long-running workflows (patent-creator, prior-art-searcher, mpep-expert, patent-drafter, etc.)
- Slash Commands - Quick-access workflows (
/create-patent,/search-prior-art,/review-claims,/full-review, etc.) - Hooks System - Custom event-driven automation
Technology Stack
Architecture:
FastMCP (MCP Server Framework)
+-- 20 MCP Tools (search, analysis, generation)
+-- Claude Code Plugin (skills, agents, commands, hooks)
RAG Pipeline:
FAISS Vector Search (BGE-base-en-v1.5, 768-dim embeddings)
+ BM25 Lexical Search (rank-bm25)
+ Cross-Encoder Reranking (MS-MARCO MiniLM-L-6-v2)
+ HyDE Query Expansion (optional, API-based)
Data Sources:
- MPEP (Manual of Patent Examining Procedure)
- 35 USC (Patent Statutes)
- 37 CFR (Patent Regulations)
- Subsequent Publications (USPTO updates)
- BigQuery patents-public-data (76M+ patents)
ML Stack:
PyTorch 2.9+ (CUDA 12.8 for GPU acceleration)
Sentence Transformers 5.1+
HuggingFace Transformers 4.57+
FAISS 1.12+ (CPU/GPU)
Validation & Monitoring:
Pydantic v2 (type safety + input validation)
Structured logging (JSON/human formats)
Performance tracking (@track_performance)
Health check system
Installation
Option 1: One-Line Install (Recommended)
# Installs package, detects GPU, downloads MPEP, builds index, registers MCP server
pip install git+https://github.com/RobThePCGuy/Claude-Patent-Creator.git && patent-creator setup
# Restart Claude Code after completion
What happens automatically:
- Installs Python package dependencies
- Detects hardware (NVIDIA GPU/Apple Silicon/CPU)
- Uninstalls CPU-only PyTorch if GPU detected
- Installs correct PyTorch (CUDA 12.8/MPS/CPU)
- Restarts setup with GPU-enabled PyTorch
- Downloads MPEP PDFs (500MB) from USPTO
- Builds hybrid index with GPU acceleration
- Registers MCP server with Claude Code
Option 2: Manual Installation
# Clone repository
git clone https://github.com/RobThePCGuy/Claude-Patent-Creator.git
cd Claude-Patent-Creator
# Optional: Create virtual environment
python -m venv venv
source venv/bin/activate # Linux/macOS
venv\Scripts\activate # Windows
# Install package
pip install -e .
# Run setup wizard
patent-creator setup
Option 3: Claude Code Plugin (Standalone Mode)
# In Claude Code
/plugin marketplace add C:\Users\<YOUR_USER>\Desktop\Projects
/plugin install claude-patent-creator-standalone
# Run setup command
/setup-patent-system
Quick Start
Using MCP Tools
# Search MPEP for claim definiteness requirements
search_mpep("claim definiteness 35 USC 112(b)", top_k=5)
# Search 76M+ patents for prior art
search_patents_bigquery("neural network training", limit=10)
# Analyze claims for compliance
review_patent_claims(claims_text)
# Generate patent diagram
render_diagram("block", components, connections)
Using Claude Code Skills
Skills activate automatically based on your task:
"Install the system"→ setup-assistant skill"Search for patents about AI"→ patent-search skill"Review my claims for compliance"→ patent-reviewer skill"Find MPEP section on enablement"→ mpep-search skill"Build the index"→ index-manager skill"Something is broken"→ troubleshooting-assistant skill
Using Slash Commands
/create-patent # Complete patent creation workflow (55-80 min)
/search-prior-art # Prior art search with novelty analysis
/review-claims # Claims-only 35 USC 112(b) analysis
/review-specification # Specification-only 35 USC 112(a) analysis
/review-formalities # MPEP 608 formalities check
/full-review # Parallel review (claims + spec + formalities)
Using Autonomous Agents
# Long-running workflows that work independently
"Create a patent for my invention, use the patent-creator agent"
"Conduct prior art search for my invention, use the prior-art-searcher agent"
CLI Commands
# System health check
patent-creator health
# Rebuild MPEP index
patent-creator rebuild-index
# Verify MCP configuration
patent-creator verify-config
# Check BigQuery authentication
patent-creator check-bigquery
# Download MPEP PDFs only
patent-creator download-mpep
# Download all sources (MPEP + 35 USC + 37 CFR)
patent-creator download-all
Requirements
Minimum
- Python: 3.9 - 3.13 (3.14 experimental)
- RAM: 8GB
- Disk: 2GB (MPEP PDFs + index)
Optional (Recommended)
- GPU: NVIDIA GPU with CUDA 12.8 (5-10x faster indexing and search)
- Google Cloud: Project with BigQuery enabled (for patent search)
- Graphviz: System package (for diagram generation)
Python Dependencies
mcp>=1.21.0- MCP server frameworksentence-transformers>=5.1.2- Embeddingsfaiss-cpu>=1.12.0- Vector searchnumpy>=1.26.0,<2.0.0- Array operations (CRITICAL: <2.0 for FAISS compatibility)rank-bm25>=0.2.2- Lexical searchgoogle-cloud-bigquery>=3.38.0- Patent searchpydantic>=2.10.0- Validationgraphviz>=0.21- Diagram generationPyMuPDF>=1.26.0- PDF processing
See pyproject.toml for complete dependency list.
MCP Tools Reference
MPEP Search Tools (2)
search_mpep- Hybrid RAG search with filtersget_mpep_section- Retrieve full section content
Patent Search Tools (7)
check_bigquery_status- Verify BigQuery configurationsearch_patents_bigquery- Search 76M+ patentsget_patent_bigquery- Get full patent detailssearch_patents_by_cpc_bigquery- Search by CPC classificationsearch_uspto_api- USPTO API searchget_uspto_patent- Get USPTO patent detailsget_recent_uspto_patents- Recent filings
Analysis Tools (3)
review_patent_claims- 35 USC 112(b) compliancereview_specification- 35 USC 112(a) adequacycheck_formalities- MPEP 608 compliance
Diagram Tools (2)
render_diagram- Generate patent diagramsget_diagram_templates- List available templates
Prior Art Tools (1)
search_prior_art- Automated prior art discovery
System Tools (5)
get_index_stats- Index statisticscheck_diagram_tools_status- Graphviz statuscheck_patent_corpus_status- Corpus availabilitycheck_uspto_api_status- API connectivityget_patent_details- Combined patent retrieval
Configuration
Environment Variables
Create .env file in the project root:
# Required for BigQuery patent search
GOOGLE_CLOUD_PROJECT=your-project-id
# Optional API keys (for HyDE query expansion)
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
# Optional settings
PATENT_LOG_LEVEL=INFO # Logging verbosity
PATENT_LOG_FORMAT=human # Log format (json/human)
PATENT_ENABLE_METRICS=true # Performance tracking
PATENT_MPEP_USE_HYDE=false # HyDE query expansion
PATENT_MPEP_DEVICE=gpu # Device (gpu/cpu)
PATENT_OPERATION_TIMEOUT=300 # Timeout (seconds)
# Windows only (for Git Bash)
CLAUDE_CODE_GIT_BASH_PATH=C:\dev\Git\bin\bash.exe
BigQuery Setup (Optional)
# Install Google Cloud SDK
# https://cloud.google.com/sdk/docs/install
# Authenticate
gcloud auth application-default login
# Set project
export GOOGLE_CLOUD_PROJECT=your-project-id
# Test connection
patent-creator check-bigquery
Skills System
Claude automatically activates specialized skills based on your task:
| Skill | Triggers | What It Does |
|---|---|---|
| setup-assistant | Installation, configuration, authentication | Complete setup lifecycle |
| development-assistant | Adding features, creating tools | Feature development workflows |
| index-manager | Building, rebuilding, optimizing index | MPEP index management |
| troubleshooting-assistant | Errors, performance issues | 6-step diagnostic methodology |
| testing-assistant | Running tests, validation | Test suite execution |
| patent-reviewer | Reviewing applications for compliance | Expert review with compliance checking |
| patent-search | Searching patents, prior art | BigQuery + PatentsView search |
| mpep-search | Finding MPEP sections, regulations | Hybrid RAG search |
| patent-diagrams | Creating technical diagrams | Graphviz diagram generation |
| prior-art-search | Novelty/FTO analysis | Prior art discovery workflows |
Each skill includes comprehensive reference documentation in skills/[skill-name]/reference/.
Autonomous Agents
For long-running workflows requiring uninterrupted focus:
| Agent | Use Case | Duration | Output |
|---|---|---|---|
| patent-creator | Create complete USPTO-ready application | 55-80 min | Specification, claims, abstract, diagrams, validation report |
| prior-art-searcher | Comprehensive prior art search | 15-30 min | Patentability report, top 10 prior art, claim strategy, IDS list |
Usage:
"Create a patent for [invention], use the patent-creator agent"
"Search prior art for [invention], use the prior-art-searcher agent"
Agents work independently while you continue other tasks.
Architecture
claude-patent-creator/
├── mcp_server/ # MCP server core
│ ├── 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
│ │ ├── mpep_tools.py
│ │ ├── analyzer_tools.py
│ │ ├── bigquery_tools.py
│ │ ├── uspto_search_tools.py
│ │ ├── diagram_tools.py
│ │ ├── prior_art_tools.py
│ │ └── system_tools.py
│ └── index/ # FAISS + BM25 index (git-ignored)
├── skills/ # Claude Code skills (15)
│ ├── setup-assistant/
│ ├── patent-reviewer/
│ ├── patent-search/
│ └── ...
├── agents/ # Autonomous agents (10)
│ ├── patent-creator.md
│ ├── prior-art-searcher.md
│ └── ...
├── commands/ # Slash commands
│ ├── create-patent.md
│ ├── search-prior-art.md
│ └── ...
├── hooks/ # Event-driven automation
├── scripts/ # Testing and utilities
├── pdfs/ # MPEP PDFs (git-ignored)
└── CLAUDE.md # Project documentation
Performance
Index Build Times
- CPU-only: 25-35 minutes (MPEP + 35 USC + 37 CFR)
- GPU (CUDA): 3-5 minutes (5-10x faster)
- GPU (Apple Silicon MPS): 8-12 minutes (2-3x faster)
Search Performance
- MPEP Search: 50-200ms (hybrid FAISS + BM25)
- BigQuery Search: 1-3 seconds (76M+ patents)
- USPTO API: 500ms - 2s (rate-limited)
Resource Usage
- Index Size: 500MB - 1GB (FAISS + BM25)
- RAM Usage: 2-4GB (loaded index)
- GPU VRAM: 1-2GB (optional acceleration)
Known Issues & Limitations
Critical Compatibility Issues
- NumPy 2.x Incompatibility -
faiss-cpu 1.12.0is NOT compatible withnumpy>=2.0. Pin tonumpy<2.0. - PyTorch Installation Order - Install PyTorch BEFORE
sentence-transformersto avoid CPU-only installation. - Windows Path Handling - Use forward slashes in MCP config paths.
- Git Bash Required - Windows users need Git Bash for
claude mcp addcommand.
See CLAUDE.md for complete troubleshooting guide.
Current Limitations
- Project is work-in-progress, not all features fully functional
- BigQuery requires Google Cloud project with billing enabled
- GPU acceleration requires NVIDIA GPU with CUDA 12.8 (Linux/Windows) or Apple Silicon (macOS)
- Some diagram types require system Graphviz installation
- HyDE query expansion requires API keys (Anthropic/OpenAI)
Credits and Attribution
Open Source Dependencies
This project builds upon excellent open source work:
- FastMCP - MCP server framework
- FAISS - Vector similarity search (Meta AI Research)
- Sentence Transformers - Text embeddings (UKP Lab, TU Darmstadt)
- HuggingFace Transformers - ML models
- PyTorch - ML framework
- rank-bm25 - BM25 lexical search
- PyMuPDF - PDF processing
- Graphviz - Diagram generation
- Pydantic - Data validation
- Google Cloud BigQuery - Patent database access
Data Sources
- MPEP - Manual of Patent Examining Procedure, published by the USPTO
- 35 USC - United States Code Title 35 (Patents)
- 37 CFR - Code of Federal Regulations Title 37 (Patents, Trademarks, and Copyrights)
- patents-public-data - Google BigQuery public dataset containing 76M+ patents
Embedding Models
- BGE-base-en-v1.5 by BAAI (Beijing Academy of Artificial Intelligence)
- MS-MARCO MiniLM-L-6-v2 cross-encoder by Microsoft
Privacy and Security
- This repository has been checked for common PII patterns and obvious secrets
- Do not commit API keys, OAuth client secrets, private keys, or personal information
- Use environment variables or local
.envfiles to store secrets - If you find any personal information in files, please open an issue
Security Reporting
If you discover a security issue or accidental leak of secrets/PII:
- Open an issue on the project's issue tracker marked as
security - For accidentally committed secrets, remove from repo history and rotate immediately
Trademark and Attribution
- This project uses factual references to USPTO public resources (MPEP, Open Data Portal APIs) for functionality and documentation
- "USPTO" is a registered trademark of the United States Patent and Trademark Office
- This project is not affiliated with, endorsed by, or sponsored by the USPTO
License
MIT License - See LICENSE file for details.
Contributing
This project is open to contributions. Since it's a work in progress:
- Expect breaking changes
- Documentation may be incomplete or outdated
- Some features may not work as described
- Issues and PRs are welcome
Development Guide
See CLAUDE.md for:
- Complete architecture documentation
- Development workflows and patterns
- Skills and agents reference
- Testing procedures
- Troubleshooting guides
- Version compatibility matrix
Support
- Issues: GitHub Issues
- Documentation: CLAUDE.md
- Examples: See
skills/andagents/directories
Roadmap
- Complete implementation of all MCP tools
- Add evaluation metrics for RAG performance
- Support for international patent offices (EPO, WIPO)
- Web interface for non-Claude Code users
- Advanced claim dependency graph visualization
- Automated obviousness analysis (35 USC 103)
- Patent portfolio analysis tools
- Integration with patent drafting software
Built with Claude Code - This project demonstrates the full capabilities of the Claude Code plugin system, including skills, autonomous agents, slash commands, and MCP server integration.