I'm evaluating Claude Code for some internal SRE tooling (think custom Prometheus exporters, Grafana data source plugins). The performance is impressive, but before I let it near our production codebase, I need to understand the data flow.
My main questions are:
* What code snippets, prompts, and generated outputs are transmitted to Anthropic's servers?
* Is there any local processing, or is everything sent for remote execution?
* How is this data retained, used for model training, or potentially reviewed?
From my initial scan of the docs, it seems like the standard API data usage policies apply, but I'm looking for the concrete, technical specifics relevant to a coding assistant. For example, if I paste a config containing internal hostnames or an alert rule with a snippet of our proprietary query logic, where does that land?
I'd love to see a breakdown similar to a network trace. Something like:
**Likely Transmitted:**
* The entire prompt context (my instructions, pasted code)
* The generated code completions
* Metadata (language, tool usage events)
**Likely Local:**
* The actual execution of shell commands, `kubectl`, etc. (this seems to run on my machine)
Can anyone who's dug deeper confirm or correct this? A clear privacy model is a must-have before I can advocate for its use in a sensitive environment.
- away