Is /security-review enough on its own?
Anthropic's security review asks Claude to look for vulnerabilities in your code, and it can find logic flaws that rule-based tools miss. But AI reviews are not deterministic, and in a 2025 test by Checkmarx the review dismissed a real remote-code-execution bug as a false positive and accepted dangerous code because a comment called it safe test code.
| AI review (/security-review) | Rule-based scanners (AppLedger) | |
|---|---|---|
| Same answer every run | No | Yes |
| Can be talked out of a finding by a comment | Yes | No |
| Known vulnerable package versions | Sometimes | Yes, from vulnerability databases |
| Leaked keys and tokens | Sometimes | Yes, dozens of known key formats |
| Logic flaws (who may see what) | Sometimes | Rarely |
What AppLedger checks
Secrets
API keys, tokens and passwords written into the code (Trivy).
Vulnerable packages
Dependencies with published security holes, and the version that fixes them (Trivy).
Risky code
SQL and command injection, unsafe data loading, debug mode left on (Opengrep).
Plain English + fix
Each finding says what could happen and what to do, plus a fix request for the agent.
Add it to Claude Code in one command
Needs uv. Without it: pip install appledger, then python -m appledger path/to/app. The full plugin with a guided “check my app” skill is on GitHub.
Does it work with Cursor and Codex?
Yes. It's a standard MCP server, so any agent that supports MCP can call it, and the command-line tool works anywhere Python runs. The fix request is plain Markdown that any coding agent can follow.
What happens to my code?
By default only the files git would commit are zipped (no dependencies, media, files over 1 MB or secret files such as .env and keys), sent over HTTPS to our scanning service after you confirm, checked, and deleted right after. With --local and the scanners installed, your code never leaves your computer. See the privacy policy.
Whether your authorisation logic is right, business-logic flaws, and settings outside the code such as database access rules or cloud permissions. That's where an AI review or a human review still earns its keep.
Questions
Does it replace /security-review?
No. Use both. /security-review can reason about logic; AppLedger's rule-based scanners give the same answer every time and can't be talked out of a finding.
Does it work with Cursor and Codex?
Yes. It's a standard MCP server and a command-line tool, so any agent that supports MCP or can run a command can use it.
Is it free?
Yes. The check, the plain-English report and the fix request are free. The package is MIT-licensed.
What happens to my code?
In the default hosted mode, source files are sent over HTTPS, scanned with Trivy and Opengrep, and deleted right after. In local mode your code never leaves your computer.