Skip to content

claude-code.md

Latest commit

 

History

History
40 lines (27 loc) · 1.48 KB

claude-code.md

File metadata and controls

40 lines (27 loc) · 1.48 KB
Raw

Claude Code Adapter

AgentWatch provides a first-class CLI wrapper for Claude Code, allowing you to observe its reasoning steps, tool calls, and apply safety policies without modifying any code.

Usage

Simply prefix your claude command with agentwatch watch:

agentwatch watch "Build me a React component for a weather dashboard"

How It Works

  1. Subprocess Interception: AgentWatch runs Claude Code as a child process.
  2. Stream Parsing: It captures the terminal output and parses reasoning traces in real-time.
  3. Safety Enforcement: If Claude Code attempts a dangerous command (e.g., rm -rf /), AgentWatch identifies the intent and can block the execution based on your policy.
  4. Dashboard Integration: Every step is streamed over WebSockets to your local AgentWatch dashboard.

Configuration Options

Pinning a Model

By default, AgentWatch respects the model Claude Code is configured to use. You can override this for auditing purposes:

agentwatch watch "..." --model claude-3-5-sonnet-20241022

Applying a Safety Policy

Use a custom policy defined in your AgentWatch server:

agentwatch watch "..." --policy strict-production

Benefits for Claude Code Users

  • Auditability: See exactly what Claude did while you were away from the terminal.
  • Safety: Prevent accidental file deletions or credential leaks.
  • Cost Tracking: Monitor the token cost of complex coding tasks.