I built a Python MCP server that lets Claude Code inspect real production systems

### What my project does I’ve been hacking on an open source project written mostly in Python that exposes production systems (k8s, logs, metrics, CI, cloud APIs) as MCP tools. The idea is simple: instead of pasting logs into prompts, let the model call Python functions that actually query your infra. Right now I’m using it with Claude Code, but the MCP server itself is just Python and runs locally. --- ### Why Python Python ended up being the right choice because most of the work is: * calling infra APIs * filtering noisy data before it ever hits an LLM * enforcing safety rules (read-only by default, dry-run for mutations) * gluing together lots of different systems The “AI” part is honestly the least interesting bit. Most of the complexity lives in normal Python code. --- ### Who this is for People who: * write Python for infra / DevOps / SRE stuff * are curious about MCP servers or tool-based agent backends * don’t want autonomous agents touching prod Not a toy project. I’ve been using earlier versions during real incidents. --- ### How it’s different This isn’t a prompt wrapper or an agent framework. It’s just a Python service with explicit tools. If the model can’t call a tool, it can’t do the thing. --- Repo (Python code lives here): [https://github.com/incidentfox/incidentfox/tree/main/local/claude_code_pack](https://github.com/incidentfox/incidentfox/tree/main/local/claude_code_pack) Happy to answer questions about the Python side if anyone’s curious.