Check your setup
Confirm whether Codex or Claude Code is detected on your machine.
npx --yes linmas@0.9.1 detectOpen source. Defensive by design.
Build with AI.
Ship with confidence.
Give your AI coding agent a security review toolkit. Find risks, understand the impact, and get concrete fixes before your next deploy.
Explore 11 focused security skills →
Apache-2.0 licensedHuman review required
review this API change before I deployHigh severitySQL injection
User input reaches a raw database query.
- Impact
- An attacker may read or modify database records.
- Decision
- Fix before deploy
- Fix
- Use a parameterized query and validate input length.
- Verify
- Add a regression test, then re-run the review.
Illustrative output · not a recorded model run. You review the evidence and make the final call.
v0.9.1current release
11security skills
7native MCP tools
3policy packs
Apache-2.0license
Codex-firstportable design
HOW IT WORKS
From “it works” to
“I’ve reviewed it.”
Four clear steps. No security jargon required. Your agent helps with the review; you stay in control of the decision.
Choose a change
Review a named diff, file, endpoint, feature, or deployment plan.
Use the right skill
Start with the router when you do not know which security domain applies.
Get an actionable finding
See severity, risk, evidence, concrete remediation, and validation steps.
Fix and verify
Apply the remediation, add a regression test, and re-run the review.
THE RIGHT SKILL FOR THE JOB
Small, focused skills.
A stronger review.
Start with the domain router when the risk is unclear. Use a specialist when you already know the type of review you need.
Explore all 11 skills →Build securely
Code, architecture, and cloud changes
Operate securely
Production readiness, monitoring, and incidents
Investigate and validate
Authorized defensive analysis
Route and govern
Choose the right workflow and document controls
LET’S GET STARTED
Your first review starts here.
New to security? Start with the Secure Code Reviewer first. It is the simplest checkpoint for code created through a vibe coding workflow.
Install one focused skill
Start small with the Secure Code Reviewer instead of installing everything blindly.
Review one real change
Give the agent a named file, diff, endpoint, or feature and ask for remediation plus verification.
Recommended for beginners
Install one security skill
Add the Secure Code Reviewer to a detected Codex or Claude Code managed skill directory.
npx --yes linmas@0.9.1 install linmas-secure-code-reviewerRequires Node.js 24+. Run
npx --yes linmas@0.9.1 detect first and npx --yes linmas@0.9.1 doctor afterward.Verification and troubleshooting.Try this after installation
Review the authentication and authorization changes in src/routes/users.js. Report severity, risk, concrete fixes, and how I should verify each fix. Do not modify files yet.
Codex primary path
Install the Codex plugin
Add the public GitHub marketplace and pin the current release.
codex plugin marketplace add TanKimGwan/linmas --ref v0.9.1
codex plugin add linmas@linmas
codex plugin marketplace list --json
codex plugin list --jsonRestart Codex completely and create a fresh task after installation.Already installed? Follow the pinned-version upgrade guide ↗
No model call
Run the offline proof demo
Inspect the review pipeline without provider credentials or network model execution.
git clone --branch v0.9.1 --depth 1 https://github.com/TanKimGwan/linmas.git
cd linmas
npm ci
npm run demo:judgeThe default demo replays a checked-in synthetic SQL-injection finding and validates the result in memory.
You build it.
We help you review it.
Security may not be your first concern when you are learning to turn prompts into working products. Linmas adds a practical review checkpoint without needing a dedicated security team.
Find your starting point below.
Vibe coders
You built a working product with AI and need a clear answer before putting it online.
Solo developers
You move quickly but do not have an AppSec specialist reviewing every change.
Engineering teams
You need consistent findings, severity, remediation, and verification across projects.
Security practitioners
You want bounded, inspectable workflows that do not overclaim automated authority.
Working code can still be unsafe code.
AI makes it easier to build features quickly. It does not automatically add authorization checks, safe database queries, secret handling, or secure deployment configuration.
review.input / common failure modes4 signals
The page works, but anyone can access another user’s data
Linmas checks ownership and authorization boundaries.
IDORThe search feature works, but user input reaches raw SQL
Linmas identifies injection risk and shows a parameterized fix.
InjectionThe app deploys, but a token or secret is exposed
Linmas prioritizes secret leakage and sensitive logging before polish.
ExposureA scanner reports a warning, but you do not know what to do next
Linmas connects each finding to risk, remediation, and verification.
No actionWorking output is not approval. Review the boundary, apply the fix, and verify again.
Keep evidence of what was reviewed, not just a chat response.
A Review Capsule is a versioned JSON record that ties one explicit input to the selected skill, normalized finding, policy decision, and the requirement for human review.
Exact input identityA SHA-256 digest and byte length connect the review to a concrete artifact.
Bounded metadataCredentials, raw provider responses, sessions, email addresses, and request identifiers are excluded.
Honest limitationA capsule is evidence of a review, not certification that software is secure.
Key fields from the v0.9.1 offline fixture. This excerpt omits required fields; use the complete example for validation. No model was called.
review-capsule.jsonOffline excerpt
{
"schemaVersion": 1,
"kind": "linmas-review-capsule",
"input": {
"source": "examples/build-week/insecure-query.diff",
"sha256": "0fec2d0a7edfa70a3edebfbbdf346c3931021b62bfe931014eebb192c068b3f2",
"bytes": 340
},
"execution": {
"mode": "offline-fixture"
},
"review": {
"specialist": "secure-code-reviewer"
},
"policy": {
"result": { "decision": "blocked" }
},
"safetyBoundary": {
"humanReviewRequired": true
}
}Know what each integration supports.
Installing a skill does not verify how an agent runs it. Native integration, managed installation, and manual reuse have different evidence and limits.
Primary / native
Codex
Git marketplace plugin, 11 skills, 7 native MCP tools, offline judge, and provider-backed review.
MCP stdio is verified on Linux with Node.js 24+. Installation and tool discovery in your host still need verification.
Managed installation
Claude Code
Managed installation of 11 skills. Claude API review is a separate integration and does not verify Claude Code runtime behavior.
Partial / format only
Hermes Agent
Skill Markdown is a candidate for manual loading. Linmas execution, MCP registration, and safety enforcement remain unverified.
Portable / manual
Gemini CLI and other agents
Reuse Markdown instructions where equivalent support exists. Linmas has no Gemini-specific installer, provider adapter, or MCP registration.
Windows supports installation and offline commands; destructive uninstall and cleanup are unavailable. Native Windows MCP and successful live provider execution are not currently verified.Read compatibility boundaries ↗
Open source. Open to inspection.
Trust should come from readable source code, repeatable validation, clear limitations, and a public contribution path. Not from marketing claims.
Snapshot verified against the Linmas main branch and npm metadata on.
Current release
Documentation and branding update: refreshed README, website logo, and version badges. No runtime changes or data migration; existing 0.9.0 installations keep the same functionality.
View release notes ↗Canonical skills
The filesystem and validation command are the source of truth for inventory.
Inspect skills ↗Policy packs
Built-in packs are schema-validated and remain advisory, with human review required.
Inspect policies ↗Open-source license
Review the license and trademark boundaries before redistribution.
Read license ↗Vulnerability reporting
Report security issues through the project’s documented process.
Read policy ↗Public implementation record
Review the dated evidence, reproducibility commands, and known limitations.
Read record ↗linmas / latest notes
Latest Articles
Practical notes on defensive security, AI-assisted development, and repeatable review workflows.
Tutorial: Test an AI Agent Approval Gate with Node.js
Build a local Node.js approval gate for AI agent actions. Test changed messages, expired approvals, and replay attempts using only built-in modules and fake data.
Read article: Tutorial: Test an AI Agent Approval Gate with Node.js
How to Build an AI Agent Security Test Plan
Build an AI agent security test plan around permissions and observable effects. Cover prompt injection, revoked access, retries, and legitimate task success.
Read article: How to Build an AI Agent Security Test Plan
A Practical Guide to Least Privilege for AI Agent Tools
Limit AI agent tools by task, identity, destination, and duration. Use this guide to separate read, draft, and execution permissions and verify the boundaries.
Read article: A Practical Guide to Least Privilege for AI Agent ToolsGOOD TO KNOW
A little clarity before you start.
Linmas is designed to lower the entry barrier, not hide its limitations. These answers describe what the project can and cannot do.
Do I need to understand application security first?
No. Start with a concrete file, diff, endpoint, or feature. Linmas should explain the risk, why it matters, the fix, and how to verify it. You still need to read the result and decide whether the evidence is sufficient.
Which skill should a vibe coder install first?
Start with linmas-secure-code-reviewer for code changes. Uselinmas-security-domain-router when you are unsure whether the problem belongs to code, cloud, architecture, incident response, or another security domain.
Does Linmas guarantee that my application is secure?
No. A review can miss context, logic flaws, deployment conditions, or new attack techniques. Linmas provides a first defensive checkpoint and structured evidence; it is not certification or automatic approval.
Does Linmas upload my code?
Offline commands do not call a model. Provider-backed reviews send the selected input after explicit confirmation. A read-only sandbox does not guarantee that the provider can read only that input. Review provider settings and accessible workspace context before sending sensitive code. Writing a proof bundle requires separate confirmation.
Is Linmas a hacking or exploitation tool?
No. Linmas is defensive and intended for owned or explicitly authorized systems. Its validation workflows emphasize bounded scope, evidence, remediation priority, and human review.
What should I do when a finding says “High” or “Blocked”?
Do not deploy the affected change yet. Confirm the evidence, apply the recommended remediation, add a regression test where possible, and repeat the review. Escalate to an experienced reviewer when the impact is unclear or affects authentication, authorization, secrets, payment, or sensitive data.
A first defensive checkpoint, not a final authority.
Linmas is designed to improve human security decisions without pretending that an AI response can certify software as secure.
What Linmas is designed to do
Provide focused, repeatable, inspectable defensive review workflows for authorized software work.
- Explain findings in practical developer language.
- Prioritize severity, risk, and remediation.
- Provide verification and regression steps.
- Keep human judgment visible and required.
What Linmas does not claim
Linmas does not replace accountable reviewers, guarantee complete isolation, or prove that software is secure.
- Not a certification or approval authority.
- Not an offensive exploitation platform.
- Not affiliated with any government institution.
- Not endorsed by OpenAI.
Your next deploy deserves a security review.
Start with the Secure Code Reviewer. Ask for evidence-linked findings, concrete fixes, and a verification step, without needing to become a security specialist first.