Tools for AI assistants to work with research administration data: an MCP server and a set of skills that let any MCP-capable AI assistant (Claude, ChatGPT, or others) answer questions about research administration data. The data is exposed through the AI4RA Unified Data Model (UDM).
This repo does not include its own agent or chat interface. Users bring the assistant they already use, and this project provides the two pieces that assistant needs to work well with research administration data: a standard way to reach the data (the MCP server) and the domain knowledge to use it correctly (the skills).
The tools are designed for Cornell's Research Admin Support System (RASS) and the PI Dashboard, but are written so other universities can adopt them without reshaping their data to match Cornell's. The UDM is the portability layer: an institution maps its local systems to the UDM once, and the same server and skills work against that common structure.
RASS and the PI Dashboard are the main repositories of research data for Cornell's Ithaca system. Their users range from high-level research administration staff to hands-on researchers, and getting answers today often means navigating dashboards or asking someone who knows where to look. General-purpose AI assistants can make that data easier to reach, but only if they can connect to it and understand how it is structured. The UDM has its own conventions, such as lifecycle-stage rows, provenance columns, and an Award/Subaward attachment pattern, that an assistant will get wrong without guidance.
The project is part of The Fellowship of AI4RA, a Cornell sub-award under NSF Grant #2427549 (University of Idaho prime). It is developed by Cornell's Research Administration Information Services (RAIS) with Cornell MEng Student Team Projects participants.
- UDM MCP server: exposes research administration data in UDM form to any MCP-capable client, with read access first and no write operations.
- Skills: instructions and reference material that teach an assistant how to use the MCP server and the UDM well, including UDM structure and conventions, how to translate common research administration questions into queries, how to interpret and cite results, and what the assistant should not attempt.
- Mock UDM: a synthetic UDM database generated from the UDM's machine-readable schema, used for development, testing, and demos without real data.
Because the interface is whatever assistant the user chooses, the server follows the open MCP standard and avoids client-specific behavior. The skills are written as portable instructions and are tested against more than one assistant.
- Build a mock UDM from the UDM's machine-readable schema and integrate RASS data against it.
- Build an MCP server that exposes UDM data to AI assistants.
- Write skills that teach assistants the UDM and how to use the MCP server accurately.
- Test the server and skills across multiple assistants, and document what differs between them.
- Measure cost, speed, and output quality.
- Contribute the MCP server, skills, and lessons back to the AI4RA community as open source.
In scope are the UDM MCP server, the skills, the mock UDM and mapping artifacts, evaluation scripts and test sets, and design documentation. Out of scope are building a custom agent loop or chat interface, document processing workflows (a separate track, see ra-doc-workflows), production deployment, and Weill Cornell Medicine data, which sits outside the Ithaca system and uses a separate eRA system.
This repo contains no real institutional data. Only synthetic or fully de-identified sample data may be committed. Work against real RASS data happens only in approved Cornell environments, with the access and security reviews that requires
GPL 3.0 - See LICENSE for details.
See the contributors list.