The idea would be to add an MCP server for CIME that exposes structured data describing models, machines, compsets, grids, case options, and other configuration metadata.
This would be a first step toward going from an experiment expressed in natural language to an actual CIME case. Instead of an AI agent relying on documentation or its own understanding of CIME, it could query the MCP server and retrieve structured data representing the capabilities and constraints of the current environment. This provides a grounded source of truth and helps avoid invalid configurations or unsupported combinations.
We could additionally provide tools to create, modify, build, submit, and monitor cases. AI agents are already capable of generating the commands needed to perform these tasks, but they are often prone to hallucinating options, workflows, or environment-specific details. Exposing these operations through MCP tools would provide a reliable interface for interacting with CIME while reducing the amount of guesswork required by the agent.
The combination of structured data and executable tools would allow AI agents to reason about CIME using authoritative information rather than inferred knowledge. This makes it possible to validate experiment configurations, guide users toward supported options, and translate high-level experiment descriptions into runnable cases with a much higher degree of confidence.
The idea would be to add an MCP server for CIME that exposes structured data describing models, machines, compsets, grids, case options, and other configuration metadata.
This would be a first step toward going from an experiment expressed in natural language to an actual CIME case. Instead of an AI agent relying on documentation or its own understanding of CIME, it could query the MCP server and retrieve structured data representing the capabilities and constraints of the current environment. This provides a grounded source of truth and helps avoid invalid configurations or unsupported combinations.
We could additionally provide tools to create, modify, build, submit, and monitor cases. AI agents are already capable of generating the commands needed to perform these tasks, but they are often prone to hallucinating options, workflows, or environment-specific details. Exposing these operations through MCP tools would provide a reliable interface for interacting with CIME while reducing the amount of guesswork required by the agent.
The combination of structured data and executable tools would allow AI agents to reason about CIME using authoritative information rather than inferred knowledge. This makes it possible to validate experiment configurations, guide users toward supported options, and translate high-level experiment descriptions into runnable cases with a much higher degree of confidence.