quickhdsdc/AAS_LLMAgent_MCP_Server

Using Asset Administration Shell as a MCP Server for LLM-Agent

The AAS MCP Server currently provides two tools for interacting with Asset Administration Shell (AAS) data.
aas_explore gives an overview on existing AAS instances and their submodels on a given AAS server endpoint.
aas_parser extracts all the submodel elements and their path from a given AAS id hosted on the AAS server endpoint and saves them in a dataframe csv for the further usage.

1. aas_explore(endpoint)

Fetches metadata about all available AAS instances from a specified AAS server.

• Parameter:
  • endpoint (string): The URL of the target AAS server.

---

2. aas_parser(endpoint, id)

Downloads the full AAS package in AASX format for a given AAS instance and extracts its property paths into a structured table.

• Parameters:
  • endpoint (string): The URL of the AAS server.
  • id (string): The unique identifier of the AAS instance.

Install

uv venv --python 3.12

In a windows cmd

.venv\Scripts\activate 
uv pip install -r requirements.txt

run and inspect MCP server with AAS tools

using stodio or sse connection

python run_mcp_server.py --transport stdio

If If using 'stdio', can directly start inspector by

npx @modelcontextprotocol/inspector python run_mcp_server.py --transport stdio

If using 'sse' transport mode, should start the MCP server at first. Then start MCP Inspector. The default MCP endport is at "http://localhost:8000/sse"

npx @modelcontextprotocol/inspector python run_mcp_server.py --transport sse

Our Work on AAS + LLM Integration

We explore how to conceptualize, develop, and deploy AAS Submodels with the support of Large Language Models (LLMs), particularly for enabling semantic search and interoperability in manufacturing systems.

1. Conceptualizing and Deploying AAS Submodels with LLM Support

This work demonstrates how to leverage LLMs for semantic search to support the creation and deployment of AAS Submodels, with a focus on quality control and zero-defect manufacturing.

Citation:

Dachuan Shi, Philipp Liedl, Thomas Bauernhansl
Interoperable information modelling leveraging Asset Administration Shell and Large Language Model for quality control toward zero defect manufacturing
Journal of Manufacturing Systems, Volume 77, 2024, Pages 678–696
https://doi.org/10.1016/j.jmsy.2024.10.011

---

2. Mapping Data into AAS Entities (Semantic/Entity Matching)

This study presents a dual mapping approach using fine-tuned LLMs to align unstructured or semi-structured data with AAS entities, enabling interoperable knowledge representation.

Citation:

Dachuan Shi, Olga Meyer, Michael Oberle, Thomas Bauernhansl
Dual data mapping with fine-tuned large language models and Asset Administration Shells toward interoperable knowledge representation
Robotics and Computer-Integrated Manufacturing, Volume 91, 2025, Article 102837
https://doi.org/10.1016/j.rcim.2024.102837