Mnemosyne MCP

Inspiration

Having worked on enterprise-level software solutions, we noticed that these have become extremely large, complex, and intertwined. A large number of microservices exist and need to be considered when writing new code for maximum efficiency.
AI agents do not have this context or memory, thus they write code that is short-sighted, duplicative, and does not leverage existing services.

So we thought — if agents like Kiro could be given the ability to understand the cluster and the microservices within it, they could become extremely powerful when working with such legacy systems.

Furthermore, if Kiro could actually look at the architecture, fetch logs, resource usage, etc., it could help with debugging errors and outages observed. If we integrate with PagerDuty, Kiro could become the perpetual on-call engineer, ensuring high reliability and uptime.

What it does

Mnemosyne is an MCP server that integrates directly into the Kiro IDE. It is deployed within the cluster itself and provides a direct way for Kiro to interact with and understand the server beyond the standard CLI tools.

We attach a serviceDescription to the pod, which is read by Mnemosyne. The service description file is itself generated by LLM agents and acts as both a human- and machine-readable document. We save these in a central repository to allow larger context-based analysis by LLMs and humans alike. These are attached to the pod during the deployment pipeline or by leveraging Kiro’s Agent Hooks.

At will, Kiro can use Mnemosyne to read this service description, instantly learning about a particular service instead of having to analyze any of the code itself.

It improves both Spec coding and Vibe coding:

Spec coding is enhanced by integrating cluster knowledge into the design specs themselves, ensuring efficient decisions and considerations.
Vibe coding is improved by allowing quick debugging of issues seen in the cluster and enabling new engineers to learn in-depth information about the cluster by simply asking Kiro to explain workflows.

These context-aware agents are much more efficient, both when producing code and debugging issues. We have opened paths to integrate with further technologies such as deployment pipelines and PagerDuty alerting.
These agents can also now potentially be employed to perform complex tasks such as end-to-end testing and penetration testing.

How we built it

We say this project is "With Kiro, For Kiro, By Kiro."

We set up a basic MCP server in Python using existing packages and then expanded it with Kiro, integrating more tools for better functionality.
Leveraging both Spec and Vibe coding modes helped us stay focused on the core issues while being efficient.
Kiro acted as a test dummy, allowing us to instantly test our code. Agent hooks further helped us create a test environment on the fly to simulate real services while keeping the main server code deployed.
With just a few words, Kiro also enabled us to integrate much-needed code quality items like tests, health endpoints, and metrics.

Challenges we ran into

The Python package works only on stdin/stdout, but we wanted to communicate via HTTP to allow the server to sit in the cluster. We had to develop an MCP-HTTP bridge to make this possible.

Accomplishments we're proud of

Apart from the MCP-HTTP bridge, we are proud of:

Plug-and-play ability.
Seamless integration with Kiro and how quickly Kiro adapted to best use Mnemosyne MCP to achieve its goals.
Intuitive language processing — we can just ask, “Describe the xyz workflow.”

What we learned

This was our first time working with an MCP server, and the limitless possibilities it opens are exciting and inspiring.
Further, working with Kiro and its various features taught us to leverage LLMs to a much better extent. We learned how to communicate with LLMs more efficiently, improving our own efficiency as a result.