Vercel’s research found that skills invoke at a stunning 0% rate when the syntax is off by even a single field. One wrong name format — poof, your entire skill is invisible. I’ve been there: spent an afternoon debugging why Claude wasn’t picking up a carefully crafted review skill, only to find a trailing space in a YAML key. Hours gone because of a whitespace character.

So meet agnix — a 423-rule AI linter and LSP built specifically for agent configs. It checks CLAUDE.md, SKILL.md, hooks, MCP configs, and more across every major coding assistant. Think of it as eslint for your agent toolchain, but way more focused.

What Makes agnix Different

Now most linters check your code. But agnix checks the rules you give your AI — harder to debug when that breaks. And it validates against official specs from Claude Code, Codex CLI, Cursor, GitHub Copilot, and more. That’s 10+ tools covered under one hood.

But here’s what got me: a config that works perfectly in Cursor can silently fail in Claude Code because they expect different formats. Still, no errors. No warnings. Just a skill that never fires. But agnix catches those cross-tool mismatches before you waste an afternoon wondering why nothing changed. If you’re running multiple assistants, you’ll want to pair this with the Agent Skills quick review — between config validation and structured commands, your agents actually do what you tell them.

Quick Start: I Ran It on My Project

So I pointed agnix at my own monorepo on the Ryzen 9 — what I thought were well-maintained agent files:

$ npx agnix .

The output hit me:

CLAUDE.md:15:1 warning: Generic instruction 'Be helpful and accurate' [fixable]
  help: Remove generic instructions. Claude already knows this.

.claude/skills/review/SKILL.md:3:1 error: Invalid name 'Review-Code' [fixable]
  help: Use lowercase letters and hyphens only (e.g., 'code-review')

Found 1 error, 1 warning
  2 issues are automatically fixable

So in a project I thought was clean, I had two issues. And the skill name Review-Code with a capital letter? But completely invisible to Claude. I was writing skills that were never executed. That’s the silent-killer scenario agnix is built for.

And the install is dead simple:

# npm (works everywhere)
npm install -g agnix

# Homebrew on macOS/Linux
brew tap agent-sh/agnix && brew install agnix

Supported Tool Coverage

I checked this against their docs page — the Claude Code rules went from 47 to 53 between last week and this release. So they’re actively iterating.

Editor Integration: LSP-Level Diagnostics

Still, agnix isn’t just a CLI tool. It runs as an LSP in VS Code, JetBrains, Neovim, and Zed. So you get inline diagnostics as you edit your config files — red squiggles on invalid fields before you even save. No context switching.

And the auto-fix mode (--fix) handles HIGH and MEDIUM confidence fixes automatically. --fix-safe only touches HIGH confidence changes. Still not sure? There’s even a browser-based playground where you can paste any config file and see diagnostics instantly. I tested it — pasted my broken SKILL.md and got the same errors the CLI found, rendered in the browser in about 2 seconds.

What to Watch Out For

So here’s my honest take: agnix is still early. v0.32.0 just dropped, and with 978 commits and 283 stars, it’s active but not massive. But the Claude Code rule set (53 rules) could still grow — some edge cases around hook scripts aren’t fully covered yet. And if you use a niche coding assistant that’s not on the support list, you’re out of luck.

But the core value is already undeniable. If you use more than one AI coding tool — which is basically everyone I know — you can bet you’re running misconfigured agents right now. agnix finds them in seconds.

Bottom Line

agnix fills a gap nobody else has touched: linting the configs that control your AI tools. It’s essential if you use 2+ coding assistants, and eye-opening even with just one. So here’s my verdict — run npx agnix . on your project. But I’ll bet you find at least one issue you didn’t know about.