convez.ai
AgenteDevOps & InfraestruturaIniciante
agentlint

AgentLint

agentlint/agentlint

Audite repos em 30s para readiness de agentes IA com score e fix prompts

Ver no GitHub
BomPluginScore · de 0 a 100

Ferramenta de auditoria que verifica se um repositório está pronto para agentes de código (Claude Code, Cursor, Copilot, Gemini CLI). Gera um score de 0-100 em 30 segundos e fornece um plano de correção automático com prompts predefinidos para integração com agentes.

Instalação

$npx convez-ai agentlint/agentlint

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
3
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 3 dias
Versão
v2.3.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Preparar repositórios para uso com Claude Code ou Cursor
  • Verificar conformidade com convenções de agentes (CLAUDE.md, AGENTS.md, etc)
  • Gerar relatórios estruturados (JSON/Markdown) para CI/CD
  • Automatizar correções via prompts de remediação predefinidos

Veredito

AgentLint resolve um problema real: alinha repositórios com os requisitos de múltiplos agentes de código via auditoria rápida e prompts de remediação. Excelente para times que usam múltiplas ferramentas de IA e querem padronização automática.

README

A tradução para o português deste README ainda está na fila.

agentlint

The Lighthouse for AI coding agents. Audit any repo for how ready it is for Claude Code, Cursor, Codex, Copilot, and Gemini CLI. Get a 0–100 score and a fix list in 30 seconds.

Quick start

npx @agentlinthq/cli

That's it. No install, no signup. Drop the command into any repo and you'll get a colored report in your terminal plus a full HTML report on disk.

# Scan the current directory
npx @agentlinthq/cli

# Scan a different path
npx @agentlinthq/cli ./packages/api

# Machine-readable output for CI or for AI agents to parse
npx @agentlinthq/cli --json > report.json

# Markdown report for AI agents to consume directly
npx @agentlinthq/cli --markdown

# Also audit the docs site
npx @agentlinthq/cli --url https://docs.example.com

# Print a copy-paste fix prompt for your AI agent (no AI involved — predefined templates)
npx @agentlinthq/cli prompt

Or install once and use the short command:

npm i -g @agentlinthq/cli
agentlint

End-users invoke it via npx/pnpm dlx/bunx regardless of what package manager their project uses — the published @agentlinthq/cli package is universal.

Note on the name. The unscoped agentlint name on npm is held by an unrelated package, and the matching org name was not available either, so the npm org is agentlinthq. The CLI is published as @agentlinthq/cli; the binary it installs is still agentlint. See ADR-0011.

Why this exists

Every AI coding tool now reads project context from a config file:

ToolFile it reads
Claude CodeCLAUDE.md (or AGENTS.md)
OpenAI Codex CLIAGENTS.md
Cursor.cursor/rules/
GitHub Copilot.github/copilot-instructions.md
Gemini CLIGEMINI.md (or AGENTS.md)

If you don't have any of these, the agent guesses. It guesses badly. The fix is simple: write the right files. agentlint tells you which ones are missing and what's in them.

What it checks

Five categories, 30 checks, 0–100 score. See AGENTS.md for project conventions.

For agent integration: agentlint --json and agentlint --markdown produce structured reports. Tell Claude Code or Cursor "run agentlint and fix what's failing" and it'll do the work.

Fix prompts (agentlint prompt)

agentlint doesn't just diagnose — it hands you the remediation. Every failed check carries a predefined fix prompt (a static template parametrized with your detected package manager, language, and workspaces — agentlint never calls an AI). Get them three ways:

# One consolidated prompt covering every finding, ordered by score impact.
# Paste it into Claude Code, Cursor, Copilot, Codex, or any coding agent.
agentlint prompt

# Just one rule
agentlint prompt --rule agents-md-exists

# Straight to the clipboard (macOS)
agentlint prompt | pbcopy

The prompts also appear per-finding in the --markdown and --json outputs (fix.prompt) and as collapsible blocks in the HTML report.

Agent skill

This repo ships a ready-made skill at skills/agentlint/SKILL.md that teaches any agent the audit → fix → verify workflow (using the installed agentlint binary, or npx -y @agentlinthq/cli@latest when not installed). For Claude Code:

cp -r skills/agentlint ~/.claude/skills/agentlint

Then ask: "audit this repo's agent readiness and fix the findings".

Optional: push reports to a dashboard (--push)

Off by default — the CLI is local-first and never phones home unless you opt in.

If you want a history of runs at https://agentlint.sh/dashboard:

# Generate a token at https://agentlint.sh/dashboard/tokens
export AGENTLINT_TOKEN=agl_...

agentlint --push
# → Pushed: https://agentlint.sh/dashboard

The token is read from AGENTLINT_TOKEN first, then from ~/.config/agentlint/token (one line, chmod 600 recommended). The token is never read from argv. Push failures (auth, rate limit, network) print one line and exit 0 — the local audit still succeeds. The CLI refuses to send to non-HTTPS URLs (except localhost / 127.0.0.1).

PR comments via the agentlint GitHub App

When a --push run is associated with a pull request and the agentlint GitHub App is installed on the repo, the App posts a single comment on the PR with the score, a diff versus the previous run, and links to the dashboard and badge. Subsequent pushes update the same comment instead of stacking new ones. PR detection works automatically on GitHub Actions pull_request events; for other CI vendors set AGENTLINT_PR=<n> or pass --pr <n>.

See packages/cli/README.md for the full security model.

How this project is built

agentlint is agent-built, open to everyone.

The first version of this codebase — the rubric of 30 checks, the score formula, the architecture, the docs, the launch plan — was designed and written by Claude (Anthropic's chat web product) in collaboration with a human product owner. Ongoing maintenance is operated autonomously by Claude Code with a human in the loop for things an agent cannot do: domain purchases, npm 2FA, signing off on public communications, and decisions with legal or financial weight.

This is not a stunt. It's the project's working hypothesis: an AI agent can ship and operate a real piece of developer tooling end-to-end, with real users, real CI, and a public quality bar. agentlint is the test of that hypothesis. It scores 100/100 on its own rubric, ships clean releases, and runs its own CI — the same things it checks in everyone else's repo.

The full operating model is public:

This project is open under MIT to human and agent contributors equally. Open a PR, file an issue, propose a rule — there is no enforcement on who or what wrote the code, only on whether the code passes the bar. The agent reviews and merges contributions like any maintainer.

Development

This repo uses pnpm workspaces. Install pnpm first: npm i -g pnpm or corepack enable.

pnpm install
pnpm run ci          # build + typecheck + lint + test + self-audit
pnpm run agentlint . # run agentlint on this repo (eat our own dog food)

Commits in this repo are co-authored by Claude via a prepare-commit-msg hook. See CHARTER.md § 6 for the rationale.

License

MIT.

Ver o original no GitHub

Bom para

Relacionados