MCP - Model Context Protocol server for OPA

[TangoIndiaPapa]

Hi, any plan or anyone writing official MCP server for OPA?
I'm starting to investigate MCP server for policy enforcement and don't want to reinvent the wheel if someone is already working on it.
I just started to investigate the following url: https://github.com/modelcontextprotocol/python-sdk

My plan is to have a "Policy MCP Server" that uses OPA to check before it allows further action.
Currently, I'm just using a yaml file that list some key policies like "do not print PII data in log" but could imagine that a large company may need to use RAG to get entire company's policy statement or use some combination that list key policies in yaml file that gets loaded during the init and the rest is searched against RAG based on some key words.

[TangoIndiaPapa]

Initial implementation of fastMCP 2.0 for MCP server for enforcing policies.
https://github.com/TangoIndiaPapa/policy-mcp-server

It's just using hard coded policy file and haven't implemented OPA yet.

[anderseknert]

Hi there! I know of a few projects that use OPA for authorization in MCP servers. I've toyed around some with an MCP implementation in Regal, which in addition to eval, could also answer questions about potential issues in a policy. I never pushed that code though as other things got in the way. Don't be afraid to reinvent the wheel! That's how we improve the wheel :)

[TangoIndiaPapa]

Thanks! Much appreciate the encouraging words!

[charlieegan3]

Hi there @TangoIndiaPapa, thanks for taking the time to look into this. I am not aware of any ongoing work on this topic.

OPA is typically used where the results of policy evaluation must be verifiably correct. For example, if someone is granted access to something they should not have been granted access to, we need to be able to review that decision in future and understand why it happened. This is a requirement of various legal and compliance requirements OPA users commonly face.

That said, I can see that this something interesting for decisions where this is not required, and where performance of external API calls is unproblematic.

It might be something we could feature on https://www.openpolicyagent.org/ecosystem when you've had more time to look into the project.

[TangoIndiaPapa]

Thanks! You are touching on the key issue for AI system! I'm trying to find a way to "review how AI came to a particular decision" to generate code or logic. Obviously, some initial input prompt along with subsequent prompts or adjustments would have impacted its decision. But, prompt engineering alone can't solve code provenance or Explanability AI(XI) issue. We need a deterministic and repeatable mechanism to enforce its behavior. Thus, I came upon OPA! :) Looks like https://play.openpolicyagent.org/ has new AI samples and it was helpful! Thank you for your help. Here is non-production work I'm working on. It's bear minimum integration using docker running OPA with Policy MCP Server. https://github.com/TangoIndiaPapa/policy-mcp-server/tree/feature/opa-integration

I'm just learning OPA rules so it's not quite doing what I want. But, the idea is to have a list of disallowed algorithm or logic so when user asks for it, it will deny it. We don't want agentic system to write "anonymous function" all over the place for the sake of efficiency....