AI代码守护

Installation options

For platform-specific details, use the setup guides in docs/:

• Claude Code
• Cursor
• Codex
• Gemini
• Kimi
• Grok
• Continue.dev
• OpenAI Agents SDK
• MCP hosts
• Generic integrations

Quick start

Pick the AI coding agent you actually use:

pip install agentlint
cd your-project
agentlint setup claude    # Claude Code
agentlint setup cursor    # Cursor
agentlint setup codex     # Codex CLI
agentlint setup gemini    # Gemini CLI

Run agentlint setup --help for the full platform list. agentlint setup resolves the absolute path to the binary, so hooks work regardless of your shell's PATH — whether you installed via pip, pipx, uv, poetry, or a virtual environment.

When AgentLint blocks a dangerous action, the agent sees:

⛔ [no-secrets] Possible secret token detected (prefix 'sk_live_')
💡 Use environment variables instead of hard-coded secrets.

The agent's action is blocked before it can write the secret into your codebase.

The setup command: - Installs hooks or guardrails for the selected agent platform - Creates agentlint.yml with auto-detected settings (if it doesn't exist)

To remove AgentLint hooks:

agentlint uninstall

Configuration

Create agentlint.yml in your project root (or run agentlint init):

```yaml

Optional for self-hosted/local API:

export AGENTCHUTE_API_URL=http://localhost:8000/v1 ```

When enabled, AgentLint sends privacy-safe event summaries only: file paths and lengths for writes/edits, truncated Bash command previews, truncated prompt previews, violation metadata, and rule counts. It never sends raw file contents, full edit strings, or full prompts.

Merge with existing config

agentlint import-agents-md --merge ```

When AGENTS.md exists and stack: auto is set, AgentLint also uses it for pack auto-detection.

Diagnose common misconfigurations (including custom rules validation)

agentlint doctor

CLI Integration

Run any command-line tool as a PostToolUse check. AgentLint executes the command after Write/Edit and reports non-zero exit codes as violations:

rules:
  cli-integration:
    commands:
      - name: ruff
        on: ["Write", "Edit"]
        glob: "**/*.py"
        command: "ruff check {file.path} --output-format=concise"
        timeout: 10
        severity: warning

      - name: pip-audit
        on: ["Write", "Edit"]
        glob: "**/requirements*.txt"
        command: "pip-audit -r {file.path}"
        timeout: 30
        severity: warning

      - name: pytest-related
        on: ["Write", "Edit"]
        glob: "src/**/*.py"
        command: "pytest tests/ -k {file.stem} -x -q --tb=short"
        timeout: 60
        severity: info

Claude Code marketplace plugin

Claude Code users can also install the marketplace wrapper from mauhpr/agentlint-plugin. The plugin repo contains Claude-specific marketplace metadata, hook files, and plugin commands; this repo contains the core engine and cross-platform setup.

AGENTS.md integration

AgentLint supports the AGENTS.md industry standard. Import conventions from your project's AGENTS.md into AgentLint config:

```bash

Scan changed files for CI pipelines

agentlint ci --diff origin/main...HEAD ```

.github/workflows/agentlint.yml

name: AgentLint on: [pull_request] jobs: lint: runs-on: ubuntu-latest steps: - uses: actions/checkout@v4 with: { fetch-depth: 0 } - run: pip install agentlint - run: agentlint ci --diff origin/${{ github.base_ref }}...HEAD ```

Only ERROR violations fail the build. Warnings are reported but don't block. Use --format json for machine-readable output.

Comparison with alternatives

FAQ

Does AgentLint slow down my AI coding agent? No. Rules evaluate in <10ms. AgentLint runs locally as a subprocess — no network calls, no API dependencies.

What if a rule is too strict for my project? Disable it in agentlint.yml: rules: { no-secrets: { enabled: false } }. Or switch to severity: relaxed to downgrade warnings to informational. The circuit breaker also helps — if a rule fires 3+ times in a session, it automatically degrades from blocking to advisory.

Is my code sent anywhere? No. AgentLint is fully offline by default. It reads the local hook, guardrail, MCP, or webhook payload and evaluates rules locally. No telemetry, no network requests. AgentChute sync is a separate opt-in path and sends only privacy-safe event summaries.

Can I use AgentLint outside Claude Code? Yes. AgentLint supports real-time blocking hooks on Claude Code, Cursor, Kimi, Grok, Gemini, Codex, and Continue.dev. For OpenAI Agents SDK and MCP hosts, use guardrail-based integration. The CLI also works standalone in any CI pipeline.