Patent Connector: MCP server for AI-powered patent research

I've built Patent Connector, a Model Context Protocol server now in open beta. It connects ChatGPT Desktop, Claude Desktop, and other MCP-compatible AI tools directly to patent databases - starting with the free EPO OPS API.

Instead of manually searching patent databases, copying claims, and pasting them into your AI chat, you can simply ask. Patent Connector handles the data retrieval automatically, turning patent research into a natural conversation.

OPS data is extracted from the EPO's bibliographic, worldwide legal event, full-text and image databases and  is therefore from the same sources as the Espacenet and European Patent Register data.

The problem with web search

Tools like ChatGPT can access Google Patents through web search, but that approach is far from optimized. AI models tend to confuse dates, misinterpret publication versus filing dates, or extract incomplete claim text from HTML. Direct API integration provides structured, validated data that the AI can process reliably.

Beyond single-provider access

The multi-provider architecture offers practical advantages. Privacy-conscious users can route requests through EPO or other providers.
More importantly, you can combine data from multiple sources in a single conversation. The system isn't limited to retrieval either. Future provider integrations will support management operations - save patents to collections or track legal status changes. The MCP protocol handles both read and write operations equally well.

Technical implementation

Patent Connector eliminates the need to copy patent data between your AI chat and web browsers. Natural conversations where the AI retrieves patent data in real-time - claims, prior art searches, family analysis - with API calls handled automatically through the MCP protocol.

I designed the architecture with encrypted credential storage and personal access tokens for MCP clients. A web UI manages provider credentials and token generation. The system currently supports EPO OPS and USPTO

Try it: patent.dev/patent-connector
Status: free EPO OPS & USPTO access

Hosted vs on-premises deployment

I wanted to provide an easy-to-setup MCP version for users who prefer not to manage API connections and servers themselves - probably most of us. Therefore, I opted to provide a hosted version of this MCP.

For law firms, running the MCP Server on-premises makes sense. While it's essentially a forwarder to APIs, privacy and confidentiality considerations matter. An on-premises deployment ensures that only the necessary external requests go directly to patent office servers like EPO or USPTO, with no third-party intermediaries handling your queries.

For patent retrieval, you could even connect your own on-premises patent database to the adapter. While this might seem like overkill for most use cases, consider this scenario: I recently set up an on-premises AI server hosting a RAG system based on AnythingLLM, which supports MCP server connections like Patent Connector.

Combining these systems, an IP law firm can deploy a chatbot that searches through client PDFs, quotes relevant sections, and incorporates data from patent databases - all without leaving the company network except for communication with patent office servers. You could ask the chatbot to retrieve a client's patent specification, explain it, and find related prior art resources, all while maintaining complete data privacy.

If you're interested in running Patent Connector locally on-premises, get in touch: Wolfgang Stark - info@patent.dev

The state of MCP

MCP is clearly in its early stages, and if you browse Hacker News discussions, you'll find no shortage of criticism about authentication complexity and spec issues. My experience building Patent Connector tells a different story. Implementing OAuth wasn't the nightmare some describe - even integrating it with this blog's existing authentication system worked without major friction. Yes, there are rough edges and the ecosystem needs time to mature, but the core concept is solid. I did observe connection issues with Claude Desktop while Claude web worked flawlessly - raising the question of whether we're dealing with MCP limitations or implementation quality in specific clients. Direct API access for AI tools solves real problems that web search and manual copy-paste workflows can't address efficiently. I'm optimistic about where this protocol is headed.

Working on patent data integration or need guidance on building robust API clients? I specialize in this exact intersection of patent information systems and modern software architecture.
Wolfgang Stark - info@patent.dev