A ModelContextProtocol Server(mcp) for Scoring with SAS Viya

6 Likes

If you are a SAS programmer who wants to part of the agentic-ai world this article is for you.

MCP servers based on the Model Context Protocol (MCP) from Anthropic is one of the popular additions to the agentic-ai world.

Users can mix and match capabilities supplied by multiple vendors from the comfort of a client like GitHub Copilot
Experts can build MCP servers that is guaranteed to work on any MCP enabled client

The brilliance of MCP is its simplicity. Since its introduction in November 2024 its adoption has been nothing short of amazing. You can find many of them listed here.

So what does this mean for SAS developers and users?

SAS developers create models using products like Model Studio and by writing custom SAS code that takes macro variables as input. Typically this is followed by developing some UI for the end users to access these models. The MCP server approach provides a nice alternative to deliver these models to the user.

Terminology:

In the rest of this article, "model" refers to all of these:

Models created with SAS solutions like Model Studio, Intelligent Decisioning etc...
User written SAS programs
Functions that call any SAS products using REST API

Details on the scoring MCP server

This article describes the viya-scoring-mcp-server to score models described above. The source code is this repository It is provided under the Apache-2.0 license.

This diagram shows the flow from a user prompt to the final response.

MCP Server in Action

(view in My Videos)

MCP server capabilities

The MCP server has the following tools for scoring.

Simple tool

devascore - calculates a special score given two numbers. This is useful for verifying that the server is functioning. Also an easy to follow example of a tool

Data related tools

findLibrary - check if specified library exists
listLibrary - list available libraries in cas or sas
findTable - check if specified table exists in specified library in cas or sas
listTables - list tables in a specified library in cas or sas
readTable - read records from a cas or sas table

Scoring with Models in MAS

findModel - check if specified model exists in MAS server
listModels - list models published to MAS
modelInfo - display the input and output variables for a specified model
modelScore - score using the selected model

Scoring with SCR

scrInfo - display the input and output variables for a specified SCR instance
scrScore - score using the specified SCR instance

Scoring with SAS code

superstat - an example of accessing custom SAS code

Using the default MCP server

Follow these basic steps

Step 1: Start the MCP server

Issue this command from any shell on your desktop

npx @sassoftware/mcp-serverjs@latest

Make sure that you have a node version >=22

Step 2: Enable github copilot for the MCP server

Similar methodologies can be used with other MCP enabled copilots.
Go to the vscode settings and search for MCP. Then select Model Server Context Protocol. Edit its config json
Add the following to the list of MCP servers

{
"viya-scoring-mcp-server": {
"type": "http",
"url": "http://localhost:8080/mcp"
}
}

The name can be anything you like.

Sidebar: Authentication with your Viya Server

This MCP server cli works similar to SAS supplied sas-viya cli commands. Use the following command to create the necessary token and refresh token. You need to do this once every 90 days or whenever the refresh token expires.

Issue this command and follow instruction: sas-viya profile init
Issue this command and follow the instructions: sas-viya auth loginCode

Issuing prompts

Use the copilot prompt area to issue prompts. Here are some sample prompts

list models
sho

Notes

This demo server is "stateless" - it does not cache any values, including any Viya sessions the tools might have created.

In a production system the designer has to make decisions on what needs to be cached and the implications of such caching.

The implication of this design choice is felt most when the tool needs to create a compute session - the requests will take longer than when the compute session is cached.