Optimized for CLAUDE.md · Free & Open Source

Is your agent
sharp & secure?

Built on Anthropic's CLAUDE.md best practices. Lint your agent's clarity, structure, security, memory, and consistency in one command. Catch leaked secrets, vague instructions, and broken references before they cost you.

View Source

Free & open source · No config needed · Runs in seconds

~/my-agent

Why this matters

Your agent config is code.
Treat it like code.

A single CLAUDE.md file can dramatically change how your agent performs. Vague instructions produce vague results. Leaked secrets become vulnerabilities. Contradictions across files cause unpredictable behavior.

“Be specific: ‘Use 2-space indentation’ is better than ‘Format code properly.’”

Vague instructions fail silently

"Be helpful" gives your agent nothing to work with. Ambiguous pronouns, missing priorities, and naked conditionals degrade output quality without any error message.

40%

of agent configs score below 60 on Clarity

Secrets hide in plain text

API keys, tokens, and passwords end up in CLAUDE.md files that get committed to repos. Standard .gitignore doesn't catch secrets embedded in markdown.

1 in 5

workspaces have exposed credentials

Multi-file configs drift

SOUL.md defines one persona, CLAUDE.md another. TOOLS.md references files that don't exist. As your workspace grows, contradictions multiply.

3.2×

more contradictions in 5+ file workspaces

How it works

One command. Full diagnosis.

No setup. No config files. Point it at your workspace and get an instant, actionable report.

01

Scan

Discovers every .md file in your workspace — CLAUDE.md, SOUL.md, USER.md, TOOLS.md, rules/, skills/. Parses structure, references, and content.

Supports Claude Code, OpenClaw, Moltbot, Cursor, Windsurf, and any Agent Skills–compatible workspace.

02

Score

Evaluates across five dimensions: structure, clarity, completeness, security, and cross-file consistency. Each scored 0–100.

Rules based on Anthropic's official best practices plus community-contributed patterns from real agent workspaces.

03

Fix

Every issue comes with a prescription and a suggested fix. Secrets get flagged for rotation. Contradictions get resolved.

Each diagnostic includes a 💡 Fix suggestion. Apply them to your files and re-run to verify your score improves.

NEW in v0.7.0

Security Scanner

Scan skills before they
compromise your agent.

The MoltX incident exposed 440,000 agents to private key theft through a malicious skill. AgentLinter now scans skills for hidden attack vectors before you install them.

3-Layer Attack Structure

L1

Manifest

Metadata tricks — fake names, hidden permissions, misleading descriptions

L2

Skill File

Malicious code — remote eval, secret exfiltration, wallet drains

L3

Prompt

Injection payloads — LLM manipulation, context poisoning

CRITICAL

Remote Code Injection

curl|bash, eval(), dynamic requires

CRITICAL

Key Theft

Private key, seed phrase, wallet access

DANGEROUS

In-band Injection

Prompt manipulation, context override

CRITICAL

Forced Wallet Connect

Unauthorized transaction signing

Verdict Levels

SAFENo threats detected
SUSPICIOUSReview recommended
DANGEROUSKnown risk patterns
MALICIOUSActive threat detected
DEFAULT SCAN

Score + Skill scan + Share (all-in-one)

PRE-INSTALL CHECK

Verify external skill before installing

Smart Detection

Project vs Agent. Auto-detected.

AgentLinter automatically detects your workspace type and adjusts recommendations. No configuration needed.

PROJECT MODE

Claude Code Projects

Detected when only CLAUDE.md is present.

  • ✓ Project-scoped rules
  • ✓ No memory strategy requirements
  • ✓ No USER.md recommendations
  • ✓ No session handoff checks
AGENT MODE

OpenClaw / Moltbot Agents

Detected when AGENTS.md, openclaw.json, or moltbot.json exists.

  • ✓ Full rule set applied
  • ✓ Memory strategy checks
  • ✓ User context recommendations
  • ✓ Session handoff validation

Scoring Engine

Eight dimensions. Real rules.

Not a vibe check. Every score is backed by specific, documented rules derived from Anthropic's guidelines, security best practices, and patterns from high-performing agent workspaces.

Structure

12%
  • File organization & naming conventions
  • Section separation & hierarchy
  • Required files present (CLAUDE.md, etc.)
  • Consistent frontmatter format
🔴 CRITICAL Missing TOOLS.md — referenced in CLAUDE.md:12

Clarity

20%
  • Naked conditionals without criteria
  • Compound instructions (too many per line)
  • Ambiguous pronouns & vague language
  • Missing priority signals (P0/P1/P2)
⚠️ WARN "be helpful" → specify: response length, tone, format

Completeness

12%
  • Identity / persona defined
  • Tool documentation present
  • Boundaries & constraints set
  • Error handling & workflows
⚠️ WARN No error recovery workflow — add escalation path

Security

15%
  • API key / token / password detection
  • Injection defense instructions
  • Permission boundaries defined
  • Sensitive data handling rules
🔴 CRITICAL Secret: API key "sk-proj-..." in TOOLS.md:14

Consistency

8%
  • Cross-file reference integrity
  • Persona alignment (SOUL ↔ CLAUDE)
  • Permission conflict detection
  • Language mixing patterns (ko/en)
🔴 CRITICAL SOUL.md persona ≠ CLAUDE.md persona — reconcile

Memory

10%
  • Session handoff protocol
  • File-based persistence (daily notes, logs)
  • Task state tracking (progress files)
  • Learning loop & knowledge distillation
⚠️ WARN No handoff protocol — agent loses context between sessions

Runtime Config

13%
  • Gateway bind (loopback only)
  • Auth mode enabled (token/password)
  • Token strength (32+ chars)
  • DM/group policy restrictions
  • Plaintext secrets in config
🔴 CRITICAL Gateway bind "0.0.0.0" — exposes agent to network

Skill Safety

10%
  • Dangerous shell commands (rm -rf, curl|bash)
  • Sensitive path access (~/.ssh, ~/.env)
  • Data exfiltration patterns
  • Prompt injection vectors in skills
  • Excessive permission requests
🔴 CRITICAL Skill contains: curl ... | bash

How we're different

Anthropic built the foundation.
We built the linter.

Anthropic's Claude Code provides CLAUDE.md memory and skills — the building blocks. AgentLinter analyzes whether you're using them effectively.

Feature
Claude Code
Anthropic Official
AgentLinter
Scoring
Basic via /init
6-category (0-100) per file
Scope
Single CLAUDE.md
Full workspace (all files)
Cross-file checks
Contradiction detection
Secret scanning
Keys, tokens, passwords
Fix guidance
Prompting suggestions
Actionable fix per issue
Custom rules
.agentlinterrc per team
CI/CD
GitHub Action per PR
Templates
/init bootstrap
4 starter templates
Reports
Web report + Share on X
Frameworks
Claude Code only
CC, OpenClaw, Moltbot, Cursor, Windsurf

Not a replacement — an extension.

AgentLinter builds on Anthropic's CLAUDE.md standard and the Agent Skills open standard. Think of it as ESLint for JavaScript — the language gives you the syntax, the linter tells you if your code is good.

Reports

Your score card.
Share it.

Every run generates a web report with tier grade, category breakdown, prescriptions, and percentile ranking. Track progress over time.

Tier grades: S → A+ → A → B+ → B → C
Exact prescriptions with actionable fix suggestions
Percentile ranking against all agents
Progress tracking: watch 72 become 89
One-click share on X with Score Card image
agentlinter.com/r/a3f8k2
Score Report
A
87
Top 12%
of all agents
Structure
80
Clarity
90
Completeness
85
Security
95
Consistency
75
Memory
88
Runtime Cfg
92
Skill Safety
98
Top issues
CRITICALRotate exposed API key
WARNAdd error recovery strategy
TIP3 issues with suggested fixes

Intelligence

Rules that learn.

Every lint teaches us something. Common failures become new rules. Bad fixes get replaced. The engine improves with every run.

Lint
Share
Users
Data
Rules ↻
L1 · Auto

Weight Tuning

Scores shift based on which warnings users fix immediately vs ignore.

L2 · Semi

Rule Discovery

Patterns found in top-scoring agents become new rule candidates.

L3 · Auto

Fix Evolution

Low-acceptance fixes get replaced through A/B testing.

L4 · Semi

Template Updates

Starter templates evolve based on what files users add.

All data anonymized · opt-out: --no-telemetry

Privacy & Security

Your files never leave
your machine.

All scanning and scoring runs 100% locally. Your file contents never leave your machine. Report sharing is optional — when enabled, only scores and diagnostic messages are uploaded (not your actual files). Use --local to skip sharing entirely.

Files stay local

Local-First Execution

All scanning runs on your machine. File contents never leave. Report sharing (scores + diagnostics only) is optional — use --local to disable entirely.

16 secret patterns

Secrets Auto-Masked

When AgentLinter detects a secret (API key, token, password), it appears as [REDACTED] in diagnostics. Even in shareable reports, raw secrets are never included.

Metadata only

Reports ≠ Raw Files

Shareable reports contain only scores, file names, line numbers, and diagnostic messages. Never the original file content. Your SOUL.md stays private.

MIT License

Open Source, Auditable

Every line of code is on GitHub. No obfuscated binaries, no hidden network calls. Read it, fork it, verify it. Trust through transparency.

--no-telemetry flag

No Telemetry by Default

Unlike many dev tools, AgentLinter sends zero analytics out of the box. If you opt in to anonymous usage stats, it's aggregated counts only — never content.

Air-gap compatible

CI/CD Safe

In CI pipelines, AgentLinter only outputs scores and diagnostics to stdout. No artifacts, no uploads, no external dependencies beyond Node.js itself.

TL;DR

AgentLinter reads your files locally, scores them locally, and outputs results locally. Nothing touches a server unless you choose to share a report — and even then, only scores and diagnostic messages are included, never your actual file contents.

One command.
Try it now.

Run it in your agent workspace. Get your score in seconds. No signup. No API key. No config.

100% free & open source · Click to copy · Node.js 18+

Help us help more agents

If AgentLinter helped improve your agent setup, share it with fellow developers. Every share helps the open-source agent ecosystem grow stronger.

Acknowledgments

Skill Security Scanner was inspired by @sebayaki's MoltX security analysis — thank you for uncovering the vulnerability that protects 440K+ agents today.