Secret scanning for AI agents: rules and entropy find candidates, a model decides which are real.
Klarion MCP Server
This MCP server, io.github.0x1Adi/klarion, performs secret scanning for AI agents. It uses rules and entropy to find candidate leaks in code, then applies a model to decide which candidates are real. The result is focused on identifying actual leaked secrets rather than only potential matches.
🛠️ Key Features
Secret scanning for AI agents
Candidate detection via rules
Candidate detection via entropy
Model-based decision to confirm real secrets
🚀 Use Cases
Scanning repositories for leaked secrets relevant to AI agents
Filtering rule/entropy findings to reduce false positives
⚡ Developer Benefits
Separates “candidates” from confirmed secrets using a model
Uses rule- and entropy-based detection prior to model evaluation
⚠️ Limitations
Based on the available data, specific deployment details, supported languages, and tool count are not provided.
Klarion finds secrets in your code: API keys, passwords and tokens.
It works in two steps:
Find. Fast local checks flag anything that looks like a secret.
Check. An AI model reads each one with the code around it and decides if it is real.
Pattern scanners stop at step 1, so they also flag test values, IDs and hashes. Klarion shows
you the real leaks.
It runs where the leak happens
A Claude Code hook that blocks the write before the file exists. The agent cannot route around it,
not through the file tools and not through Bash.
A coding agent is told to hardcode an AWS key pair. Klarion's pre-tool hook blocks the write at 95% confidence, and the agent reports that it cannot land the credentials through Bash or the file tools.
On a pull request it scans only what the PR adds. It caches the AI answers, so unchanged code
costs nothing. Findings show in the job log. On a private repo without GitHub Code Security,
add upload-sarif: "false" to skip the Security tab upload. GitLab CI and every option are in
the reference.
Without the stages line the hook is off and runs only when you ask for it, with
pre-commit run --hook-stage manual klarion. The klarion hook builds Klarion with Go the
first time it runs; use id: klarion-system if klarion is already installed. Everyone who
commits needs a model set up.
json
// ~/.cline/mcp.json, or the MCP settings of any other client{"mcpServers":{"klarion":{"command":"klarion","args":["mcp"]}}}
The agent calls scan_text before it writes code and scan_file before it commits. Approve
each call rather than auto-approving the server: scan_file reads any path it is given. With no
model configured the server hands each candidate and its decision rules to that agent to judge,
so it needs no API key. Configure one and Klarion judges instead. Full steps, including the
tools and how to verify the install: llms-install.md.
Results
Klarion found about 1.7× more real secrets than gitleaks, and 89% of its alerts were real
(gitleaks: 91%).
We tested Klarion v0.3.0 on CredData: 337 real open
source repos where Samsung labeled the lines that hold secrets. These numbers cover code
outside test folders.
Klarion
gitleaks
detect-secrets
Real secrets found
35%
21%
35%
Alerts that were real
89%
91%
34%
False alarms on flask and rails
9
33
247
Good to know:
Klarion skips test fixtures and docs examples on purpose. CredData counts those as real.
Counted that way, Klarion found 9% and gitleaks 46%.
Klarion's numbers come from a sample of 612 AI checks with one model (Claude Haiku). The
likely range is 30 to 40% found and 82 to 97% real.
CredData's labels came from what other scanners found, which helps those scanners.
Step 1 runs on your machine. It looks for known key formats, password = value style
lines, credential files like .pgpass, and secrets hidden in base64 or hex. It also checks
how random a string looks.
Step 2 asks the model. Each candidate goes out with a few lines of code. The model
answers secret, false positive or uncertain. Uncertain counts as a secret.
You can see what was hidden. Run klarion scan . --show-suppressed.
Your code goes only to the model you pick. Set send_secret = false to send a masked
value instead. It is less accurate.
Limits
It needs a model. Without one, klarion scan and the commit hooks stop with an error.
The Claude Code hook still blocks provider keys such as AWS, GitHub and Stripe, and the MCP
server hands the candidates and its rules to your agent to judge.
No API key? A hosted judge with no setup is coming. Join the list.
The model can be wrong. Check what it hid with --show-suppressed.
It finds secrets, it does not fix them. Rotate any secret it finds.
Local hooks can be skipped with git commit --no-verify. Run Klarion in CI too, with
the Action. The pre-commit hook checks staged changes only, so in CI it has nothing to check.