AI Agent Policies

Policies turn your Cursor Rules or CLAUDE.md into hard guardrails which an AI Agent cannot simply ignore, or forget. They handle what to do when an agent wants to make a decision, along with other hooks-supported events. Policies can yield both decisions and guidance.

This framework supports Claude Code. Support for Cursor is in beta.

Why Policies

Automating Decisions

Rule files can be forgotten or ignored completely by LLMs. Policies are unavoidable:

if re.match(r'^terraform\s+apply(?:\s|$)', command):
    yield PolicyDecision(action=PolicyAction.DENY, reason="terraform apply is not allowed. Use `terraform plan` instead.")

if re.match(r'^terraform\s+(fmt|plan)(?:\s|$)', command):
    yield PolicyDecision(action=PolicyAction.ALLOW)

In the image below, the agent is denied by policy from running terraform apply. The agent can then make decisions on what to do next, in this case it outputs that it will attempt a terraform plan as per the policy recommendation.

Automating Guidance

Aside from denying and allowing automatically, policies can also provide guidance through Post-* events:

if comment_to_code_overlap >= 0.4:
    yield PolicyHelper.guidance(
        "Ensure comments add value beyond describing what's obvious from the code. "
        "This comment may be redundant with the code it describes."
    )
    return

In the image below, an agent makes a change which includes a comment that is detected as being potentially redundant. The guidance from the snippet is given as feedback to the change, after which the agent by itself decides to remove the redundant comment.

Usage

At DevLeaps we developed an internal policy set for AI Agents. To create your own, refer to the example server as a starting point The example server contains:

• A basic server setup demonstrating the use of policies and middleware.
• Rudimentary policies showcasing how to automatically deny, allow and provide guidance.
• Rudimentary middleware showcasing how multi-command tool use could be handled.

To run the example server:

devleaps-policy-example-server

This starts a minimal server running just our example policies.

Architecture

graph TB
    subgraph "Developer Machine"
      Editor[Claude Code / Cursor]
        Client[devleaps-policy-client]
    end

    subgraph "Policy Server"
        Server[HTTP API]
        Policies[Your policies<br/>kubectl, terraform, git, python, etc.]
    end

    Editor -->|Hooks| Client
    Client --> Server
    Server -->|Events| Policies
    Policies -->|Decision and Guidance| Server
    Server --> Client
    Client -->|Decision and Guidance| Editor

Quick Start

Installation

Install from PyPI:

pip install devleaps-agent-policies

Or for development:

git clone https://github.com/Devleaps/agent-policies.git
cd agent-policies
uv sync

Running an Example Server

devleaps-policy-example-server

The example server runs on port 8338 by default and serves endpoints for both Claude Code and Cursor.

Configure Claude Code

Run the install command to automatically configure hooks:

devleaps-policy-client install

Or manually add devleaps-policy-client to your Claude Code hooks configuration in ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ],
    "PostToolUse": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ],
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ],
    "Stop": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ],
    "SubagentStop": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ],
    "Notification": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ],
    "PreCompact": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ],
    "SessionStart": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ],
    "SessionEnd": [
      {
        "hooks": [
          {
            "matcher": "*",
            "type": "command",
            "command": "devleaps-policy-client"
          }
        ]
      }
    ]
  }
}

Configure Cursor

Run the install command to automatically configure hooks:

devleaps-policy-client install cursor

Or manually create or edit ~/.cursor/hooks.json:

{
  "version": 1,
  "hooks": {
    "beforeShellExecution": [
      { "command": "devleaps-policy-client" }
    ],
    "beforeMCPExecution": [
      { "command": "devleaps-policy-client" }
    ],
    "afterFileEdit": [
      { "command": "devleaps-policy-client" }
    ],
    "beforeReadFile": [
      { "command": "devleaps-policy-client" }
    ],
    "beforeSubmitPrompt": [
      { "command": "devleaps-policy-client" }
    ],
    "stop": [
      { "command": "devleaps-policy-client" }
    ]
  }
}

Uninstall

To remove hooks from Claude Code:

devleaps-policy-client uninstall

To remove hooks from Cursor:

devleaps-policy-client uninstall cursor

This removes all devleaps-policy-client hooks from the respective editor configuration files.

Sessions

Each Claude Code or Cursor session receives a unique session_id. Policies can use this to track context across multiple hook events within the same session, enabling stateful policy decisions. See the session state utility to store and retrieve per-session data.

Configuration

The client supports centralized configuration via JSON files:

Home-level config (~/.agent-policies/config.json):

{
  "bundles": ["python", "git"],
  "editor": "claude-code",
  "server_url": "http://localhost:8338",
  "default_policy_behavior": "ask"
}

Project-level config (.agent-policies/config.json):

{
  "bundles": ["terraform"]
}

Configuration is merged with project settings overriding home settings.

Available fields:

• bundles: List of enabled policy bundles (default: [])
• editor: Editor name, ignored by client (default: claude-code)
• server_url: Policy server endpoint (default: http://localhost:8338)

Policy Bundles

Policies can be organized into bundles to group related rules for specific workflows or project types. This allows you to compose different policy sets without having to manage separate server configurations.

How bundles work:

• Universal policies (registered with bundle=None) are always enforced
• Bundle-specific policies are only enforced when enabled in config
• Multiple bundles can be enabled: set "bundles": ["bundle1", "bundle2"] in config
• Bundles can coordinate through shared session state

See the uv example for a working one-rule bundle implementation.

Development

This project is built with uv.