# MCP Manager Skill

**Intelligent on-demand MCP server loading to minimize context consumption**

## Quick Start

The skill operates automatically - Claude detects when you need MCPs and offers to load them.

### Manual Commands

```bash
# Check what's currently loaded
python3 ~/.claude/skills/mcp-manager/mcp_control.py status

# Load specific MCP
python3 ~/.claude/skills/mcp-manager/mcp_control.py enable n8n-mcp

# Unload specific MCP
python3 ~/.claude/skills/mcp-manager/mcp_control.py disable n8n-mcp

# See what MCPs would help with a task
python3 ~/.claude/skills/mcp-manager/mcp_control.py detect "your query here"

# Reset to minimal (unload all)
python3 ~/.claude/skills/mcp-manager/mcp_control.py reset
```

Or just ask:
- "Load the n8n MCP"
- "Unload all MCPs"
- "What MCPs are currently loaded?"

## Available MCPs

| MCP | Category | Description | Est. Tokens |
|-----|----------|-------------|-------------|
| `cognitive-memory` | Memory | Cognitive memory with decay scoring | Always loaded |
| `n8n-mcp` | Automation | n8n workflow management | ~1500 |
| `playwright` | Automation | Browser automation and testing | ~1000 |

## Configuration

- **Global MCPs** (always-on): `~/.claude.json` → `mcpServers`
- **Project MCPs** (on-demand): `<project-root>/.mcp.json` → `mcpServers`
- **Full Registry**: `~/.claude/.mcp-full.json` (all available definitions)
- **Logs**: `~/.claude/logs/mcp-manager.log`

> **Important:** `~/.claude/.mcp.json` is NOT read by Claude Code. Global servers go in `~/.claude.json`.

## Workflow

1. **Minimal Start** - Begin each session with no on-demand MCPs loaded
2. **Auto-Detect** - Claude detects when MCPs are needed via trigger keywords
3. **Load on Demand** - Enable only required MCPs
4. **Restart** - MCP changes require Claude Code restart
5. **Complete Task** - Use the MCP for your work
6. **Unload** - Free context by removing MCPs when done

---

**Version**: 1.1.0
**Updated**: 2026-03-05