CCO-MCP (Claude Code Oversight)

Real-time audit and approval system for Claude Code tool calls. Get instant visibility and control over AI agent actions with a sleek web dashboard.

Overview

CCO-MCP provides a security layer between Claude Code and your system, enabling you to monitor and control AI tool calls in real-time. It acts as a firewall for AI actions, allowing you to review sensitive operations before they execute while auto-approving safe ones based on configurable rules. This let's you scale to many more non-interactive Claude Code instances running in the background without incurring "tab-switching fatigue".

Related MCP server: PearClaw

Get Started

Clone and run with Docker Compose

git clone https://github.com/onegrep/cco-mcp.git
cd cco-mcp
docker-compose up

Configure Claude Code

Install CCO-MCP at the user scope:

claude mcp add -s user -t http cco-mcp http://localhost:8660/mcp

This will add the following to your Claude Code config:

{
  "mcpServers": {
    "cco-mcp": {
      "type": "http",
      "url": "http://localhost:8660/mcp"
    }
  }
}

For self-hosted deployments, replace http://localhost:8660/mcp with your deployment URL.

Visit http://localhost:8660 to access the dashboard.

Enable Approval Prompts

To use CCO-MCP for approval prompts, run Claude Code in non-interactive mode with the appropriate Permission Prompt Tool flag:

claude code -p "your prompt here" --permission-prompt-tool mcp__cco-mcp__approval_prompt

This enables Claude to request approval through CCO-MCP before executing sensitive operations without using the "dangerously skip permissions" option.

Main Features

🛡️ Smart Approval Rules

Create rules to auto-approve safe operations (like file reads) while requiring manual approval for sensitive ones (like running bash commands).

📊 Real-Time Dashboard

Monitor all tool calls as they happen with live updates via Server-Sent Events.

🔍 Detailed Audit Logs

Every tool call is logged with full context including agent identity, parameters, and approval status.

Configuration Options

All configuration can be managed through the web interface at http://localhost:8660/config.

Default Action and Timeouts

• Default Action: Choose whether unmatched requests are automatically approved or require manual review

• Auto-Deny Timeout: Set how long to wait for manual approval before automatically denying (default: 5 minutes)

• Entry TTL: Configure how long audit log entries are retained (default: 24 hours)

Approval Rules

Rules use a priority system (lower numbers = higher priority) to determine actions:

• Tool Matching: Match specific tools (built-in like Read/Write or MCP server tools)

• Agent Matching: Create rules for specific agent identities

• Pattern Matching: Use wildcards for flexible rule creation

• Actions: Set rules to auto-approve or auto-deny matching requests

Environment Variables

• PORT - Server port (default: 8660)

• CCO_CONFIG_PATH - Config file location (default: ~/.cco-mcp/config.json)

Development

Project Structure

cco-mcp/
├── src/              # TypeScript backend
│   ├── audit/        # Core audit service
│   ├── routes/       # REST API endpoints
│   └── server.ts     # MCP server
├── ui/               # React frontend
│   └── src/
│       ├── components/
│       └── pages/
└── dist/             # Build output

Design Principles

• Simplicity first - Enabling core use-case of approving background agents

• Real-time by default - SSE for instant updates

• Type safety - Full TypeScript coverage

• User-friendly - Intuitive UI with helpful defaults

Local Development

# Install required tools
just brew

# Setup project dependencies
just setup

# Run both backend and UI
just dev-all

# Run tests
just test

# Format code
just format

# Build everything
just build-all

Contributing

We welcome contributions! Please see our Contributing Guide for details.

License

MIT License - Copyright 2025 OneGrep, Inc.

Acknowledgments

Built for use with Claude Code and the Model Context Protocol by Anthropic.