Claudewatch by Blackwell Systems is a free Model Context Protocol (MCP) server tool designed for the mcp category, specifically monitoring AI agent behavior during Claude Code sessions. It functions as an "AgentOps" tool for the AI agent itself during development, providing self-awareness and post-session analytics. Claudewatch is explicitly not a memory MCP replacement, infrastructure observability solution, or LLM API monitoring tool. It operates by reading local session data from the ~/.claude/ directory, focusing on enhancing the AI development experience.

What does Claudewatch connect to?

Claudewatch connects to and monitors AI agent behavior within Claude Code sessions, serving as an AgentOps tool for the AI agent during its development lifecycle. It operates by reading local session data directly from the ~/.claude/ directory. As an MCP server tool, it supports transport protocols compatible with the Model Context Protocol, including Streamable HTTP and SSE for direct API integration. For local connections, Claude Desktop communicates with Claudewatch over stdio, while remote HTTP servers require a proxy like mcp-remote. Claude Code is the primary compatible AI client.

Core features and tools

Claudewatch exposes 29 Model Context Protocol (MCP) tools, enabling fine-grained monitoring and control of AI agent behavior. Specific callable tools include get_project_health for project status, get_drift_signal for detecting unproductive patterns, get_task_history for reviewing past actions, get_blockers for identifying known solutions to errors, and get_agent_performance for analyzing agent efficacy. These tools bring AgentOps capabilities directly into the AI development experience, offering insights into agent lifecycles and decision-making.

Who is it for?

Claudewatch is for AI agents, which gain self-awareness through its monitoring capabilities, and for developers, who receive valuable post-session analytics. It integrates AgentOps directly into the development workflow for Claude Code users. This tool helps developers understand and improve AI agent behavior by providing critical operational insights. Skip this if you require a memory MCP replacement or a general infrastructure observability solution.

How to get started

No credentials are required for Claudewatch's core function, as it operates with a "zero network calls" authentication model, reading only local files and writing to a local SQLite database. For remote MCP server connections, the Model Context Protocol supports OAuth authentication using Bearer tokens.

Use cases

• Error loops: Detects consecutive errors during Claude Code sessions and prompts Claude to call get_blockers() to identify known solutions.
• Drift detection: Identifies unproductive exploration patterns, such as "8 consecutive reads with zero writes," and alerts Claude to potential inefficiencies.
• Agent performance: Analyzes agent lifecycles, computes success rates, and provides get_agent_performance() for Claude to query its own efficacy and avoid spawning agents likely to fail.

Claudewatch provides essential AgentOps capabilities for developers working with Claude Code, offering deep insights into AI agent behavior. Its focus on local session data and a rich set of MCP tools enhances the development experience by enabling self-awareness for agents and detailed post-session analytics for developers. It supports critical workflows like error loop detection, drift analysis, and agent performance monitoring. Try monitoring a Claude Code session to detect unproductive drift using get_drift_signal.