Hey everyone! I wanted to share a CLI tool I recently created called droidctx, which is designed to enhance the capabilities of coding agents like Claude Code when investigating production environments. The core idea is to pre-generate context from various production tools, store it in Markdown format, and reference it in a file for the agent to use while debugging production issues. This approach has shown to speed up the process, reduce token consumption, and often lead to better answers from the agent.
The CLI connects to a range of production tools and pulls metadata by running `droidctx sync`. It can gather information from platforms like Grafana, Datadog, Kubernetes, Postgres, AWS, and over 20 connectors, creating a tidy directory of structured Markdown files. The end result? Fewer calls to different tools, decreased chances of the agent generating hallucinations about your unique setup, and less context needed to share during troubleshooting. For example, there was an instance where the agent pinpointed a bug in a specific table column by finding an exact query, which would have been overlooked otherwise.
The tool is MIT licensed, includes pre-built connections to monitoring services, Kubernetes, databases, CI/CD, and logging systems, and runs entirely locally so your credentials remain secure. I'm curious if others have faced similar challenges with coding agents and whether the approach of generating context once for reuse seems practical, or if there are other better solutions out there. I'd love to hear your thoughts, feedback, or any missing features you see!
4 Answers
This approach to snapshotting metadata is interesting, but I think the real challenge lies in capturing the operational context that people understand. Questions like ‘Why is this process scheduled for 3 AM every Tuesday?’ or ‘What steps do you take when this dashboard alerts?’ are hard to codify. That knowledge makes a big difference between a junior engineer and a senior one who can resolve issues quickly.
Are your agents actually debugging on production servers? That sounds pretty risky!
In my case, they only have read-only roles with restricted permissions. They've been really handy during incident response. Just the other day, an agent analyzed the production logs and the codebase during an issue, and within 5 minutes, it provided a full root cause analysis, helping to resolve the incident in about 10 minutes.
I really hope they're just debugging and nothing more.
I think pre-generating context is the right move! If you set up droidctx sync to run regularly, it will keep your Markdown files updated without hammering APIs when you need to debug. A tip: consider putting a metadata header in each file with a sync timestamp and connector version, so the agent can track data freshness. For teams with rapidly changing infrastructure, implementing a change detection system could trigger selective syncs when resources are modified.
Thanks for the feedback! I really appreciate it. I'm making that change to add timestamps and connector versions now!
On that note, do you have ideas on how a CLI could monitor for changes? It seems a bit complex for a tool that's intended to be simple.
This sounds awesome! Does it include an Azure connector, like ARM?
Yes, it actually has a great Azure connector! I'm excited for you to try it out! By the way, could you explain more about what ARM is? I’d love to learn.
Agreed! I've seen positive outcomes by adding a memory feature to the agent that learns from team chat conversations, GitHub logs, and even incident reports. It involves a lot of complexity, though, and really requires organizational support because of the sensitive nature of that data. (Just to clarify, I work at a company specializing in this area, so take that for what it's worth!)