Removes secrets from tool output before your AI coding assistant reads it.
Tool output:
DATABASE_URL=postgres://admin:secret@db.example.com:5432/prod
AWS_ACCESS_KEY_ID=AKIAEXAMPLEKEYID0001
SMTP_PASS="Qf7mZp2kLw9vTx4nRb2024"
APP_NAME=myapp
What the model reads (the third line needs SMTP_PASS learned by redacted init):
DATABASE_URL=[REDACTED:database_url ...prod]
AWS_ACCESS_KEY_ID=[REDACTED:aws_access_key ...0001]
SMTP_PASS="[REDACTED:smtp_pass ...2024]"
APP_NAME=myapp
Secrets reach the model through normal work, not only through cat .env:
- Ordinary output carries them: a committed key in
git diff, a token in a log, a stack trace that prints the environment. - Some commands exist to print them:
heroku config,kubectl get secret -o yaml,aws secretsmanager get-secret-value. - Deny lists check the command string. An npm script, a make target or a wrapper CLI changes the string, not the output.
- Once in context, a secret stays in the transcript on disk and in every later request.
redacted does not judge commands. It scrubs the output, whatever produced it.
- Runs as a Claude Code PostToolUse hook:
redacted scrubreads each tool result on stdin. - Scanning happens on your machine. No network calls.
- A secret is replaced inline by a marker that keeps its last 4 characters.
- Clean output passes through unchanged.
curl -sSL https://raw.githubusercontent.com/svn-arv/redacted/main/install.sh | sh
redacted init
Homebrew, Cargo and pre-built binaries for Linux, macOS and Windows are in the reference.
- Learns the secrets in your
.envfiles, so the hook catches them even when no vendor pattern knows their format. - Stores a SHA-256 hash, the length and an optional shape for each value. Never the value.
- You pick the values to learn and confirm the config before anything is written.
- Registers the hook in your Claude Code settings. Safe to run again.
Full walkthrough and flags: Setup: redacted init.
| Tier | What it catches | Default |
|---|---|---|
| Vendor signatures | Keys with a known format: 46 built-in patterns covering AWS, GitHub, Stripe, OpenAI, Anthropic, database URLs, private keys and more. | on |
| Learned secrets | The values redacted init learned from your.env , matched by hash, plus rotated keys with the same prefix. |
on once init writes entries |
| Entropy heuristic | Random-looking values in KEY=value lines that no other tier knows. |
off |
Email addresses and phone numbers can be scrubbed too. Each is off until you turn it on in the config.
Every pattern and what it matches: What it detects.
- Claude Code, through the PostToolUse hook.
- OpenCode, through the plugin that
redacted init --format opencodewrites.
Other assistants, and whether their hooks can replace tool output, are listed under Other tools.
-
Learned secrets match as whole tokens. One embedded mid-token, such as in a URL path, passes unless a vendor pattern catches it.
-
A learned shape also matches a non-secret with the same prefix and similar length. Check the shape
initshows before confirming. -
With the heuristic on, a high-entropy identifier that mixes case and digits (some tool or request ids) may be redacted. The 4-character hint makes these easy to spot.
-
Every pattern scans the full output in turn, after a substring prefilter. Multi-megabyte output adds noticeable latency per call.
-
Commands typed at the Claude Code prompt with
!are not tool calls. Their output reaches the model unscrubbed. -
Reference : install methods,
init, every pattern, configuration and commands. -
Architecture : how the code fits together.
-
Benchmark : recall and precision on the test corpus.
-
Contributing : development, adding a detection rule and releasing.
MIT