Agent Guardrails

Chinese | English

agent-guardrails checks code changes made by AI coding tools before you merge them. It records what the task is allowed to change, compares that plan with the diff and test evidence, and gives the reviewer a clear result.

It does not replace Claude Code, Codex, Cursor, Gemini, or OpenCode. It works beside them as a repo-level guardrail.

Why use it

AI coding tools are fast, but the hard part is often the handoff:

• What exactly was the agent supposed to change?
• Did it touch files outside the task?
• Were tests or validation actually run?
• Is there evidence a reviewer can trust later?
• Did a small request turn into a broad rewrite?

agent-guardrails makes those questions explicit and repeatable.

What it checks

• Scope: flags changes outside the declared task, allowed paths, or intended files.
• Validation: checks reported commands and required evidence files.
• Consistency: warns when a task spreads across too many files or directories.
• Risk: surfaces protected paths, interface changes, config changes, migration changes, and secret-like patterns.
• Reviewer output: prints a score, verdict, findings, next actions, and a concise review summary.
• Agent setup: writes repo-local helper files and MCP configuration guidance for supported agents.

Supported agents

agent-guardrails can generate or update helper files for:

Agent	Helper location
Claude Code	CLAUDE.md
Codex	.codex/instructions.md
Cursor	.cursor/rules/agent-guardrails-enforce.mdc
Gemini CLI	GEMINI.md
OpenCode	.opencode/rules/agent-guardrails-enforce.md

Requirements

• Node.js 18+
• Git
• A git repository for the project you want to protect

The npm package includes native runtime binaries for Windows x64, macOS x64/arm64, and Linux x64. It also includes the Node runtime.

Quick start

npm install -g agent-guardrails

cd your-repo
agent-guardrails setup . --agent codex --lang en
agent-guardrails enforce --all --lang en
agent-guardrails doctor --lang en

Use the agent name you actually work with: claude-code, codex, cursor, gemini, or opencode.

Core workflow

1. Set up the repo once.
2. Ask your agent to use agent-guardrails for the task.
3. Create a task brief with plan, or let an MCP-capable agent start the guarded loop.
4. Implement the smallest safe change.
5. Run validation, then run check --review before merge.

agent-guardrails plan \
  --task "Add input validation" \
  --intended-files "src/add.js,tests/add.test.js" \
  --allow-paths "src/,tests/,evidence/" \
  --required-commands "npm test" \
  --evidence "evidence/add-validation.md" \
  --lang en

npm test
agent-guardrails check --base-ref HEAD~1 --commands-run "npm test" --review --lang en

The screenshots above were generated from a real agent-guardrails run in a temporary git repository.

MCP integration

setup prints the MCP snippet for the selected agent. For Codex, the snippet looks like this:

[mcp_servers.agent-guardrails]
command = "npx"
args = ["agent-guardrails", "mcp"]

Once connected, MCP-capable agents can read repo guardrails, start a bounded implementation loop, check after edits, and finish with a review summary.

Commands

Command	Purpose
setup . --agent <name>	Initialize guardrails and agent helper files for a repo.
enforce --all	Add stronger guardrail instructions for all supported agents.
unenforce --all	Remove injected guardrail instructions.
plan --task "..."	Write a task contract before implementation.
check --review	Run a reviewer-facing guardrail check.
doctor	Diagnose repo setup and runtime availability.
generate-agents	Regenerate agent helper files.
mcp	Start the stdio MCP server.
serve	Start the local API service for integrations.
start, stop, status	Manage the local background daemon.

Configuration

setup creates .agent-guardrails/config.json. The most common settings are:

{
  "checks": {
    "scope": {
      "violationSeverity": "error",
      "violationBudget": 5
    },
    "correctness": {
      "requireCommandsReported": true,
      "requireEvidenceFiles": true
    }
  }
}

Useful options:

Setting	What it controls
checks.scope.violationSeverity	Whether scope violations are blocking errors or warnings.
checks.scope.violationBudget	How many minor scope slips can be treated as soft warnings.
checks.consistency.maxChangedFilesPerTask	File-count warning threshold for one task.
checks.correctness.requireCommandsReported	Whether validation commands must be reported.
checks.correctness.requireEvidenceFiles	Whether declared evidence files must exist.
checks.risk.requireReviewNotesForProtectedAreas	Whether protected areas need review notes.

Optional Verity receipt

Agent Guardrails works independently. When a local Verity verification receipt exists, pass it explicitly:

agent-guardrails check --verity-receipt ./path/to/receipt.json --review

Agent Guardrails invokes verity verify-receipt from PATH without a shell. An accepted receipt is recorded as verificationEvidence; it proves only Verity's local runtime verification and does not satisfy scope, security, protected-path, required-command, or evidence-file requirements. If Verity is unavailable or rejects the receipt, the check is blocked.

Evidence and docs

• CHANGELOG
• Workflows
• Troubleshooting
• Release evidence
• Executable scenarios

License

MIT