Every time you paste a log file into Cursor, or ask Claude Code to debug
a production error, you're probably leaking something you shouldn't.
API keys. Patient emails. Bearer tokens. Production AWS credentials.
They flow into prompts silently, by default, with no warning.
I spent this weekend building ContextDuty to fix that.
What it does
ContextDuty is a local-first context firewall. It sits between your files
and your AI assistant, scanning and redacting sensitive values before the
prompt leaves your machine.
Demo link - https://asciinema.org/a/uouCCzZe7UkbomNM (vibecoded on friend's laptop)
pip install contextduty
contextduty scan mylog.txt
contextduty redact --in mylog.txt --out clean.txt
It also runs as an MCP server — so Cursor, VS Code, and Claude
can intercept automatically without you doing anything manually.
Current state
This is a weekend build — rough edges, limited detector coverage
(email, API keys, AWS keys, bearer tokens, phone numbers), North
America phone format only. Works end-to-end, 53 tests passing,
published on PyPI.
What I'd love feedback on:
What detectors are you missing?
Does the policy layering model make sense for your team?
What would make this useful in your actual workflow?