Idea/Feature request: jqassistant as MCP Server

[qdrddr]

MCP Server that integrates jqassistant and provides it's as a set of tools for development MS/GitHub Copilot, Roo-Code/Cline and Claude Desktop.

example https://github.com/davidkingzyb/SCAST

[DirkMahler]

Going to evaluate the Neo4j MCP support, see https://github.com/neo4j-contrib/mcp-neo4j

[DirkMahler]

First results look quite promising, see https://www.linkedin.com/posts/dirk-mahler-837a4b5_ai-neo4j-claude-activity-7321596753223340033-xz4r?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAELWusBRxkwCkVK_HjnzDL7OBkziTwRbZU

There are some limitations:

• the application that has been analyzed (https://github.com/jqassistant-tutorials/ddd-architecture-with-jmolecules) is quite small
• the application is well structured
• the provided information is not accurate (message brokers)

Some ideas for further investigation:

• What happens if we take more complex, real world applications? How reliable are answers? We can check with some Open Source projects like BioJava or the Corona Warn App Server?
• Which difference makes the abstraction level of provided data, i.e. is the "scan" sufficient or do concepts (e.g. from the Spring plugin) improve accuracy of the results?
• Can we use this approach (or a similar one) with other LLMs than Claude? How do the results compare for similar questions?

[Bramaten]

Notes:

• the modell should be able to extract information about the used technologies and functionalities of the application correctly and completely
• compare the results of promts after the scan and after the analysis phase
• evaluate setup options with other modells and RAG

Criterias:

• correctness
• local running modell -> data security
• utilizability for large applications / data samples

[2015xli]

Hi, I am investigating how to implement an mcp/agent for my project that uses jqAssistant. I am wondering what's the best way to get the source Java file location from a :Type:Class node. Current class node has sourceFileName property that has no full path. Without the source Java file, it is difficult for the agent to do deep analysis that depends on code.

I tried to build the graph with both the source tree and the target tree, and I wrote some experimental rules to map them. But I haven't figured out a reliable way to connect them because java source files may have the same file name while with different "package" by using the package statement.

An approach is to use javasrc plugin, but I cannot find an up to date version (?), and not sure if that's a right approach.
Another approach is to use a simple java source parser to extract the package statement from the java source files, building a map from file tree to fully qualified class name, and then use that map to build graph relationships.

Any suggestions?

I've developed a clangd-graph-rag for C/C++ project with mcp/agent, where I have to use clang parser to get some important info that is missing in the clangd index. For jqassistant, I am thinking to taking the similar approach (i.e., the approach 2 above). Before I am doing that, I want to hear you guys' suggestions.

Thanks.

[DirkMahler]

• Currently there is no direct link to source files but the file name of a package and the source file name by a contained Java type can be used to reconstruct the relative file name to a source folder:

MATCH
  (p:Package)-[:CONTAINS]->(t:Type)
RETURN
  p.fileName + "/" + t.sourceFileName

• Support for source file mapping is an open item, but currently not yet planned (possibly required for integrations into GitHub/GitLab build pipelines and reporting)
• the javasrc plugin is not maintained, it has never left the prototype state

[2015xli]

Thanks @DirkMahler!

I've wrote a source file linking tool based on jQAssistant graph by parsing the source file to find its package name, and construct its fqn, and then link the :Type:File node with the same fqn value. This is just a quick & simply solution. The project is at https://github.com/2015xli/jqassistant-graph-rag .

Based on the tool, the project proceeds to normalize the graph, making it coding-agent friendly. An example MCP server and a coding agent are included. It works as expected for my code analysis/development purpose, which may not be the primary usage of jQAssistant graph.

There are lots of treasures in jQAssistant to be exposed to an agent. Will keep enhancing the project.

Cheers!

[DirkMahler]

Wow, that looks impressive - will definitely have a deeper look at this!

[DirkMahler]

Let's see if we can address some of your issues when implementing jQAssistant/jqassistant#1077.