More screenshots
Works with the AI Agents you already use
Secrets Leak Scanner: Find leaked keys before you push
**The problem** You're about to push, and something in the diff looks like an API key.
Free
Secrets Leak Scanner: Find leaked keys before you push
Example session with this skill installed
Scan the sample app configuration and documentation, keeping matched values out of the report.
- Read your context and instructions
- Compiled the secrets leak scanner
Secrets scan: references/tiered-demo
Gate: FAIL (fail on medium or higher; actionable tier). 2 files scanned, 2 findings (1 high, 1 info), 2 distinct values. Suppressed: 0 by allowlist, 0 inline.
ACTIONABLE: 1; CONTEXT: 1 (excluded from gate unless --include-context).
Values are redacted: length and SHA-256 fingerprint only.
ACTIONABLE
| Severity | Rule | Location | Context reason | Redacted value |
|---|---|---|---|---|
| high | gitlab-pat | app.conf:1:6 | … len 30, sha256:8668530705ea576b |
CONTEXT
| Severity | Rule | Location | Context reason | Redacted value |
|---|---|---|---|---|
| info (placeholder) | url-embedded-credentials | docs/guide.md:1:14 | example-or-test-path, url-placeholder-or-reference | … len 8, sha256:5e884898da280471 |
Skipped: 0 excluded, 0 binary, 0 too large, 0 unreadable, 0 symlinks.
Remediation order: revoke/rotate first, then remove from code, then purge history (see references/remediation.md).
This is real output from two made-up files. The fake GitLab token in the config file fails the check (exit code 1). The placeholder password in the docs is listed under Context and doesn't fail it. With a real leak, rotate or revoke the credential first, then remove it from the code, then look at history and logs.
Connects securely to your tools. The creator never sees your data.
About this skill
The problem You're about to push, and something in the diff looks like an API key. Or a teammate pasted a config file and you're not sure what's in it. You want to know where the credentials are without copying them into a chat, a ticket or your agent's reply.
See it before you buy: real input, real output Command (from the package folder, exit code 1, 0.033 s):
python3 -I -B scripts/scan_secrets.py references/tiered-demo
Before: A small folder: app.conf with a token-shaped value, and docs/guide.md with a placeholder password in an example URL. (The fake token is assembled at test time, so the package ships clean.)
After: lines copied from the script's output (excerpt):
# Secrets scan: `references/tiered-demo`
**Gate:** FAIL (fail on medium or higher; actionable tier). 2 files scanned, 2 findings (1 high, 1 info), 2 distinct values. Suppressed: 0 by allowlist, 0 inline.
Values are redacted: length and SHA-256 fingerprint only.
| high | gitlab-pat | `app.conf:1:6` | | `…` len 30, `sha256:8668530705ea576b` |
| info (placeholder) | url-embedded-credentials | `docs/guide.md:1:14` | example-or-test-path, url-placeholder-or-reference | `…` len 8, `sha256:5e884898da280471` |
Remediation order: revoke/rotate first, then remove from code, then purge history (see references/remediation.md).
What the numbers show
- Ran on 23 public repositories (9,682 files): 21 actionable matches, all read by hand: 13 were real development or CI credentials and 8 were false alarms. None was a production secret.
- 150 context matches are listed separately; I read 20 of them and none needed action. The whole run took about 22 seconds.
- Honest limit: one corpus, so it says nothing about what the scanner misses.
What you get
- One Python script with no dependencies. It checks 22 rules: 18 known token formats (OpenAI, Anthropic, AWS, GitHub, GitLab, Slack, Stripe, Google, SendGrid, Twilio, npm, PyPI, JWTs, passwords inside URLs and more), plus private keys, secrets in
.envfiles, high-randomness values assigned to names likepasswordorapi_key, and.envfiles that exist at all. - A Markdown or JSON report with file, line and column, the value's length and a short SHA-256 fingerprint. The value itself and the surrounding line never appear in the report.
- Two sections. "Actionable" is what most likely needs fixing. "Context" holds matches in docs, tests and examples, which you still see but which don't fail the check by default.
- An allowlist for reviewed false alarms (by fingerprint, path and rule), inline exceptions, exclude patterns, and counts of every file it skipped or match it suppressed.
- A workflow for your agent: it runs the scan, tells you what wasn't covered, walks through each match with you, and puts revoking or rotating a real credential first.
Measured on real data (full detail) I ran it on 23 public open-source repositories, 9,682 files in total. It flagged 21 actionable matches and 150 context matches. I read all 21 actionable ones: 13 were real development or CI credentials (mostly one test database password repeated in one repo) and 8 were false alarms, so roughly 2 out of 3 were worth acting on. None was a production secret.
I also read 20 of the 150 context matches and found nothing that needed action. The other 130 I didn't check. This is one corpus, so it says nothing about what the scanner misses. The whole run took about 22 seconds on one machine. Repository pins and every label are in references/benchmarks.md.
Known limitations
- 8 of the 21 actionable matches in the test above were false alarms. Treat the output as a list to review, not a verdict.
- The rules that move matches into "Context" can also move a real leak there. When you're investigating a suspected leak, run with
--include-context --fail-on infoso everything counts. - It reads the files as they are now. It doesn't scan Git history, open archives, read
.gitignore, or check whether a key still works. Short or all-letter passwords outside.envfiles can be missed. - Excludes, size limits and allowlist entries all reduce what gets checked. File paths and fingerprints in the report can still be sensitive, so don't post it publicly.
Who it's for Developers and reviewers who want a quick local check before pushing or sharing files, and a report they can paste into a review without leaking anything.
Requires Python 3.8 or newer (tested on 3.13). No packages to install, no Git needed. The script is a single file and doesn't need anything else from the package to run.
It never connects to the network and has no telemetry. It only writes a file if you ask for a JSON report or redirect the output.
Check it yourself
89 command-line test cases and 4 saved sample outputs, all re-run against the final script. From the package folder, follow references/checks.md to repeat them with plain Python. The 23-repository test above is documented in references/benchmarks.md.
Price Free, yours to use with any agent that can run Python.
Pairs well with
- GitHub Actions Security Audit: risky permissions, unpinned actions and secrets echoed to logs, with the line number for each.
- Dockerfile Linter: root users, unpinned base images and cache-busting COPY lines, each with a fix.
- Dependency License Audit: AGPL and unknown licenses in your lockfile, with an exit code for CI.
No Git history or remote repo scanning. Skips binary files and files over 2MB by default. Short or low-entropy passwords may be missed. Does not validate if keys are active or live.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 15 days ago
- Free with an Agensi account
Requires
Permissions
Uses a terminal/Python runner and file-read access to the user-selected input and optional allowlist. Reads standard input only when `-` is selected. Writes nothing by default; `--json-out FILE` explicitly overwrites that report, and caller redirection may write stdout. No network access, no telemetry, no subprocesses and nothing sent anywhere.