For over a decade, Open Targets has provided foundational data and resources to enhance our collective ability to identify and prioritise therapeutic drug targets. A core part of our mission has been to democratise access to target discovery evidence, empowering researchers across diverse backgrounds and expertise levels.

Today, we're excited to announce the initial release of the official Model Context Protocol (MCP) server for the Open Targets Platform. This development represents another step in our commitment to making Open Targets Platform data accessible to AI-powered research tools and agents, bringing systematic target discovery capabilities to Large Language Model (LLM)-based workflows while maintaining the rigour and accuracy our community expects.

To make this possible, we partnered with Anthropic, a leading AI provider in life sciences, as part of their Claude for Healthcare & Life Sciences launch. 

"With Claude for Healthcare and Life Sciences, we are bringing together leading organizations. Claude helps visionary researchers and institutions transform the way that they work and shape the future of their field, while setting new standards for responsible AI deployment," says Jonah Cool, Head of Life Sciences Partnerships at Anthropic. 

"We were excited to work with the Open Targets Platform given their historical scientific excellence in target discovery. This integration is a great example of bringing leading AI to platforms and data that are already widely used across the industry.

Together we worked to build an MCP that addresses specific problems and opportunities in target identification and prioritisation, and is both technically and scientifically rigorous. It is going to help even more scientists make use of Open Targets more quickly and efficiently than ever before. I look forward to seeing how the community makes use of this tool."

By combining Anthropic's technical expertise and computing resources with the extensive target discovery knowledge of the Open Targets consortium, we've created the first iteration of this MCP server. It’s available on Claude, and compatible with different AI providers and local models. We plan to improve the server through ongoing releases.

Why are we doing this?

Large language models and AI agents offer transformative opportunities to access scientific data, create innovative interfaces, and enhance research automation. However, maximising the quality and reliability of AI solutions requires access to up-to-date, high-quality information that keeps these systems well-informed and performing at their full potential.

Until now, the Open Targets Platform lacked an efficient, accurate interface for making our data accessible to LLMs. With MCP emerging as a standard protocol among AI providers, our dedicated Open Targets Platform MCP server can leverage our public API to provide the context needed for precise, efficient, and interpretable results.

This first release marks the beginning of a journey to democratise access to Open Targets Platform data in new ways, complementing our existing web interface, APIs, and datasets. The MCP server opens opportunities to develop AI-driven solutions that will ultimately assist our collective efforts to improve drug target identification and therapeutic hypothesis building.

The technical strategy

The official MCP uses modern frameworks like fastmcp and follows best practices to maximise cross-platform compatibility. We're designing flexible GraphQL access strategies with plans to expand to additional backends. Critically, we're building a business interpretation layer that helps agents understand pagination, assess the relevance of results, and determine the logical next steps.

To validate this approach, we developed the Karenina benchmarking framework, which systematically tests how well different AI configurations can answer research questions. The framework automates the process of posing questions, collecting answers, and evaluating their accuracy through both an API and a graphical interface. We've benchmarked 140 question-answer pairs across diverse setups—comparing local versus remote models, different agent architectures, and performance with and without our official MCP versus third-party alternatives. 

This rigorous testing revealed that generic API wrappers struggle with complex aggregations, multi-entity queries, and nuanced result interpretation— producing inefficient workflows, excessive token consumption, and higher error rates that undermine the reliability researchers need for critical target selection decisions. The MCP released today capitalises on many of the lessons learned from this benchmark. 

Together we worked to build an MCP that addresses specific problems and opportunities in target identification and prioritisation, and is both technically and scientifically rigorous.

It is going to help even more scientists make use of Open Targets more quickly and efficiently than ever before.

We are committed to ensuring open access to the Open Targets Platform and its data, while maintaining user confidentiality. All our work adheres to open science practices, using open-source code and permissive licenses to ensure broad reusability and benefit to the wider community. We develop flexible open source solutions that enable confidential use of the Open Targets Platform and the public MCP, and can be adapted for on-premises deployment or internal instancing. This will ensure researchers and organisations can refine their therapeutic hypotheses with the level of confidentiality they need.

Looking ahead

Developing the MCP extends beyond a single implementation. We're establishing a foundation that strengthens the interoperability and openness of the Open Targets data ecosystem, making biomedical knowledge more accessible across AI agents and chat interfaces. 

We're excited about this new direction and the possibilities it opens for researchers combining AI with systematic target discovery. As always, community use cases guide the evolution of our tools, and we'd be delighted to hear your thoughts on how the MCP might support your work.

Get started with the MCP, find out more in the documentation and GitHub repository, and please share your feedback on the Open Targets Community.