ScienceIT offers AI tools and services for Lab researchers under the CBorg AI Portal, providing Berkeley Lab staff with access to selected AI models that can perform a multitude of tasks. Many researchers have taken advantage of these offerings with creative applications and using new tools.
One of the tools that researchers are finding valuable is the Model Context Protocol (MCP), which can be layered onto CBorg’s large language models in order to extend the knowledge and capabilities of a model. MCP is an open-standard and open-source framework developed by Anthropic to enable seamless communication and data exchange between AI models and external systems. It acts as a universal translator, allowing AI systems to access and utilize information from diverse sources, like databases, APIs, and other tools.
Damian Guenzing is a postdoc in the Advanced Light Source’s (ALS) Coherence and Magnetism Program, as part of the Photon Science Operations team. He works on an LDRD project with scientist Sophie Morley using coherent X-rays to image complex materials.
“LLMs famously hallucinate if they use data from unreliable sources for training or are trying to close knowledge gaps. But as scientists we can harness LLMs to our own domain knowledge using trusted resources and established computational methods,” said Damian.
He used MCP to connect three modules — a NERSC API connector, a beamline-specific knowledge module, and an X-ray database and calculation module — to CBorg. “The MCP tools allowed me to leverage CBorg’s LLMs to access data and run calculations and simulations seamlessly, simultaneously accessing all the resources that I connected. For example, I can pull data from the X-ray database and do quick calculations with the Python module XrayDB, ask the LLM model on CBorg if NERSC is up and run my calculations on Perlmutter, or check if certain measurements are feasible at our beamline,” explained Damian.
Damian said that setting up the MCP modules didn’t take long, perhaps a week or so of intermittent work. He noted that if you start with existing packages, it’s even easier to add the MCP layer and integrate them into your AI workflow.
“The great thing is that MCP is modular by design, which makes it very easy to add tools and make the workflow even more powerful,” said Damian.
Contact ScienceIT to Connect CBorg to Databases, Resources Across the Lab
Andrew Schmeder, a consultant with the IT Division’s ScienceIT Department, which supports scientists at the Lab with their data and computation needs, thinks there is great potential value in increasing the use of MCP at the Lab. Many researchers would welcome easier access to Berkeley Lab databases outside their own divisions, especially if they are part of a multidisciplinary team.
“It could be a valuable tool for internal collaborations,” said Andrew. “Connecting CBorg to data and other resources at the Lab using MCP would accelerate development of AI-automated workflows.”
Andrew is planning to deploy Lab-hosted MCP servers within the CBorg API service, enabling authenticated users to access selected internal Lab resources. He is currently seeking use cases to pilot; interested researchers should contact him at awschmeder@lbl.gov.
For more information
Model Context Protocol: https://modelcontextprotocol.io/docs/getting-started/intro
CBORG: https://cborg.lbl.gov/
