PatentClaw – An Open-Source Tool for Automatic Patent Generation and Review in Research Contexts

1. What is PatentClaw

PatentClaw is an open-source auxiliary tool designed for researchers, university faculty and students, as well as patent agents. It specializes in automatically organizing LaTeX-formatted research papers into Chinese patent technical disclosure documents. The tool completes tasks such as paper parsing, invention point extraction, and draft writing through a multi-Agent collaboration system. It establishes traceable evidence mapping for content such as formulas, tables, and algorithms, avoiding the fabrication of experimental data. After generating an initial draft, the accompanying PatentReviewer module performs multiple rounds of rule-based and semantic reviews using a fixed checklist of 30 items, and executes localized modifications for items that fail the review, forming a complete closed-loop from paper parsing to the final output of the technical disclosure document.

Image source: Official article
Image source: official article

Technical positioning and domain: Belongs to the intersection of natural language processing and intellectual property automation, applied in the intelligent transformation scenario of converting research paper outcomes into patent documents, with a clear functional positioning in the field of automated patent writing.

Development background: Initiated and maintained by open-source community developer lzlfwow, the project aims to address the practical issues of high barriers to patent writing for researchers and long production cycles for technical disclosure documents. It achieves a pipeline-style transformation from research papers to patents through an Agent collaboration architecture.

Core value: Solves four core issues in the process of converting research paper outcomes into patent assets: document restructuring, technical feature extraction, evidence tracing, and quality review. It significantly shortens the cycle from paper publication to patent application and reduces communication costs between inventors and patent agents.

Technical features: Driven by a multi-Agent collaboration mechanism for paper parsing and patent draft generation, it ensures traceability of technical descriptions through evidence number mapping. It employs a versioned checklist of 30 items to maintain consistency in review standards, and supports a localized modification strategy that isolates issues one by one, balancing generation efficiency with output quality.

2. Key Features

• Paper-to-PCT Document Conversion: Automatically parses LaTeX research papers (supports single .tex files, project directories, or .zip archives), organizing the technical solutions, experimental data, and implementation details from the paper into a technical disclosure document that meets the requirements of Chinese invention patents, achieving a dual transformation of format and content from academic papers to patent documents.

• Multi-Agent Collaborative Generation: Five specialized agents are responsible for paper understanding, invention point mining, technical solution refinement, example organization, and disclosure document writing, forming a pipeline-style automated production mechanism. Each step is handled by an independent agent, ensuring the professionalism and parallel efficiency of the processing workflow.

• Evidence Mapping and Hallucination Prevention Mechanism: Parses the paper's section structure, formulas, tables, algorithm pseudocode, and figure captions to establish an evidence numbering mapping system, ensuring that every technical description in the disclosure document can be traced back to the original paper. For parameters, steps, or experimental results not disclosed in the paper, the system strictly refuses to fabricate them. When materials are insufficient, it marks the issue as a block and prompts for additional information.

• 30-Item Rule Review: The PatentReviewer module performs reviews based on a versioned checklist covering 13 dimensions, including 18 deterministic rule checks (e.g., format completeness, structural compliance) and 12 online semantic checks (e.g., technical logic consistency, sufficiency of examples), systematically identifying material risks and logical flaws in the disclosure document.

• Per-Problem Isolated Revision: Each failed review item is handled by an independent issue agent, submitting only localized patches for the corresponding field to avoid introducing new errors when modifying one issue, enabling precise revisions.

• Multi-Round Review Closure: After revisions, the full checklist is automatically re-executed, continuously addressing issues until all items pass or the preset maximum number of review rounds is reached, forming a convergent quality control loop to ensure the output document is thoroughly refined.

• Full-Traceability and Audit Support: Retains initial review reports, final review reports, item-by-item modification records, evidence source mapping, and comparison of differences before and after modifications, making the entire generation and review process auditable and meeting the regulatory requirements of the patent drafting process.

• Multi-Format Export and Dual-Mode Operation: Supports exporting disclosure documents in three formats: JSON, Markdown, and DOCX; operationally supports both local offline validation mode and online generation mode via the OpenAI Responses API. The generation and review modules can also be deployed as independent FastAPI services, facilitating integration into existing workflows.

3. How to Use

1. Environment Setup: Clone the GitHub repository to your local machine, create a virtual environment with Python 3.10 or higher, and install the two core packages, PatentGenerator and PatentReviewer, in development mode to ensure complete runtime dependencies.

2. Configure API Key: Set the OpenAI-compatible API key, Base URL, and model name in the environment variables to enable online generation mode. If you choose the local offline mode, no external API configuration is required, and the system can operate entirely in an isolated environment.

3. Upload the Paper: Input the LaTeX paper into PatentGenerator in the form of a single .tex file, project directory, or .zip archive. The system automatically parses the paper's section structure, formulas, tables, algorithms, and figure captions, establishing a mapping of evidence numbers.

4. Generate Draft: Run the latex2disclosure command. Five Agents work collaboratively to understand the paper, identify inventive points, refine the technical solution, and organize the implementation examples, automatically generating a draft of the technical disclosure and exporting it in Markdown or DOCX format.

5. Review and Revision: Pass the job.json file generated by the Generator and the original paper to PatentReviewer to perform a fixed checklist of 30 items. Each item that fails the check is handled by an independent issue Agent, which submits a local patch. The Reviewer automatically rechecks until all items pass or the maximum number of iterations is reached.

6. Obtain Final Draft and Manual Review: Export a complete product package that includes the preliminary review report, revision plan, change log, and final disclosure document. The inventor verifies the technical facts and parameters, and the patent attorney completes the legal review before formally submitting the patent application.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For research institutions, university research groups, and patent agencies, if the goal is to achieve bulk conversion of research outcomes with traceable evidence chains and maintain the data privacy benefits of local deployment, PatentClaw is more suitable due to its open-source nature and multi-Agent closed-loop review mechanism. The core value of this solution lies in keeping the paper data on-premises while obtaining complete evidence mapping and process auditing capabilities, making it ideal for scenarios with dual requirements for IP compliance and data sensitivity.

For independent inventors and startups aiming to quickly draft a U.S. provisional patent application, Idea2PatentAI's low-threshold interactive approach is more appealing. It can generate a 35-55 page application draft without requiring LaTeX formatting, but lacks a fixed review mechanism and evidence tracing capability. If the core requirement of the enterprise is prior art search and infringement risk assessment, DeepCite's solution is more aligned. PatentPal, on the other hand, is suitable for users who wish to complete the drafting of a full U.S. patent application through guided interaction. When selecting a tool, the priority should be evaluated across three dimensions: input format compatibility, review rigor, and data deployment method.

6. Editor's Summary

PatentClaw has chosen a differentiated technical approach in the field of automated patent generation for scientific research: it limits input to LaTeX papers, combining a multi-Agent collaborative architecture with an evidence mapping mechanism to create a complete closed-loop from paper parsing to disclosure document review. The core of its technological innovation lies in treating "evidence traceability" as a foundational principle in system design—by parsing sections, formulas, tables, and figure captions of the paper and establishing evidence number mappings, it constrains the correspondence between generated content and the source literature at a mechanistic level, effectively addressing the common "hallucination" issue in large language models when generating professional documents.

In terms of practical value, this tool precisely addresses the pain points of high barriers to patent writing and long production cycles for technical disclosure documents faced by researchers. Universities and research institutes generate a large number of scientific research outcomes with potential patent value every year, but the efficiency of converting papers into patents is constrained by dual barriers: document reconstruction and legal compliance. PatentClaw significantly reduces the time required to generate initial drafts through an automated pipeline, while its 30-item checklist and per-issue isolation revision mechanism ensure the stability of output quality. Its full-trace design supports auditability and traceability, providing substantial assistance in meeting the document standardization requirements during the patent examination process.

In terms of target users, this tool is most user-friendly for researchers accustomed to using LaTeX, and it is also suitable for patent agencies as an auxiliary tool for batch processing disclosure documents. The open-source nature and support for local offline operation provide important safeguards for data-sensitive scenarios. The project's future development potential depends on three directions: its compatibility with non-LaTeX input formats, the continuous iteration and expansion of the 30-item checklist under varying patent examination requirements, and the level of activity in its community ecosystem. The current version has clear functional boundaries and solid engineering implementation, laying a solid foundation for future evolution in the field of intellectual property automation.

7. Application Scenarios

• Patentification of Research Achievements: Researchers from universities and institutes can quickly convert published LaTeX papers into technical disclosure documents that meet the requirements of China's Patent Law, shortening the cycle from paper publication to patent application and improving the efficiency of intellectual property conversion for research outcomes.

• Technical Mining of Thesis Papers: Automatically identify patentable technological innovations from doctoral or master's theses, extract technical problems, solutions, and implementation examples, preventing research outcomes from being lost due to untimely patent applications.

• Conversion of Corporate R&D Documents: Technical departments in enterprises can organize internal technical reports and algorithm design documents into standardized patent disclosure materials, reducing communication costs between inventors and patent agents, and improving the quality of invention disclosures and the efficiency of patent mining.

• Assistant Review for Patent Agents: Patent agents can use a fixed checklist of 30 items and a multi-round revision mechanism to quickly identify material risks, logical flaws, and areas requiring supplementation in disclosure documents, automating repetitive review tasks and freeing up human resources for high-value patent strategy work.

• Intellectual Property Management for Research Teams: Research groups can systematically convert their annual paper results into patent assets in bulk, establishing a traceable technical evidence chain to support subsequent patent applications, technology transfer, and intellectual property operation decisions.

8. FAQ

Q: What paper formats does PatentClaw support?
A: Currently, it supports three formats: a single LaTeX file (.tex), an engineering directory, and a .zip compressed package. Papers created using other typesetting tools like Word must first be converted to LaTeX format, which falls within the current functional boundaries of the tool's input capabilities.

Q: What external dependencies are required for online semantic checks?
A: The 12 online semantic checks depend on an OpenAI-compatible API configuration. You need to set the API key, Base URL, and model name in the environment variables. If external services are inaccessible, you can choose to run in local offline mode, but the semantic check feature will not be available.

Q: Does PatentClaw fabricate experimental data that does not exist in the paper?
A: No. The system design includes explicit anti-hallucination mechanisms and strictly refuses to fabricate parameters, steps, or experimental results not disclosed in the paper. When the materials are insufficient, the system will mark the blocking points and prompt for additional content, ensuring that all generated document content can be traced back to the original paper.

Q: What dimensions does the 30-item checklist specifically cover?
A: The checklist is a versioned fixed standard that covers 13 dimensions, including 18 deterministic rule checks and 12 online semantic checks. The specific list is defined in the versioned configuration files in the project repository and is updated as the project evolves.

Q: Can the generated disclosure document be directly submitted for a patent application?
A: No. The technical disclosure document generated by the tool is a technical proposal document. It still needs to be verified by the inventor for the accuracy of technical facts and parameters, and reviewed by a patent agent for legal compliance and claim drafting before it can be officially submitted for a patent application.

Q: Can the generation and review modules be integrated into existing business systems?
A: Yes. The generation and review modules of PatentClaw can run as independent FastAPI services and be integrated with other systems via standard API interfaces. This makes it suitable for scenarios where the ability to generate patent documents needs to be embedded into existing intellectual property management workflows.

Q: Is commercial use of the project supported?
A: PatentClaw is released under an open-source license, allowing both commercial and academic use. Specific licensing terms are defined in the LICENSE file in the GitHub repository. It is recommended to confirm the relevant agreements before using it commercially.

9. Project Links

• Project Website: https://lzlfwow.github.io/PatentClaw/
• GitHub Repository: https://github.com/lzlfwow/PatentClaw