Convert any MCP server into a Claude Skill with 90% context savings.
MCP servers are great but load all tool definitions into context at startup. With 20+ tools, that's 30-50k tokens gone before Claude does any work.
This converter applies the "progressive disclosure" pattern (inspired by playwright-skill) to any MCP server:
- Startup: ~100 tokens (just metadata)
- When used: ~5k tokens (full instructions)
- Executing: 0 tokens (runs externally)
# 1. Create your MCP config file
cat > github-mcp.json << 'EOF'
{
"name": "github",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {"GITHUB_TOKEN": "your-token-here"}
}
EOF
# 2. Convert to Skill
python mcp_to_skill.py \
--mcp-config github-mcp.json \
--output-dir ./skills/github
# 3. Install dependencies
cd skills/github
pip install mcp
# 4. Copy to Claude
cp -r . ~/.claude/skills/githubDone! Claude can now use GitHub tools with minimal context.
The converter:
- Reads your MCP server config
- Generates a Skill structure with:
SKILL.md- Instructions for Claudeexecutor.py- Handles MCP calls dynamically- Config files
- Claude loads metadata only (~100 tokens)
- Full instructions load when the skill is needed
- Executor runs MCP tools outside context
Before (MCP):
20 tools = 30k tokens always loaded
Context available: 170k / 200k = 85%
After (Skills):
20 skills = 2k tokens metadata
When 1 skill active: 7k tokens
Context available: 193k / 200k = 96.5%
GitHub MCP server (8 tools):
| Metric | MCP | Skill | Savings |
|---|---|---|---|
| Idle | 8,000 tokens | 100 tokens | 98.75% |
| Active | 8,000 tokens | 5,000 tokens | 37.5% |
Any standard MCP server:
- ✅ @modelcontextprotocol/server-github
- ✅ @modelcontextprotocol/server-slack
- ✅ @modelcontextprotocol/server-filesystem
- ✅ @modelcontextprotocol/server-postgres
- ✅ Your custom MCP servers
Use this converter when:
- You have 10+ tools
- Context space is tight
- Most tools won't be used in each conversation
- Tools are independent
Stick with MCP when:
- You have 1-5 tools
- Need complex OAuth flows
- Need persistent connections
- Cross-platform compatibility critical
Best approach: Use both
- MCP for core tools
- Skills for extended toolset
pip install mcpPython 3.8+ required.
┌─────────────────────────────────────┐
│ Your MCP Config │
│ (JSON file) │
└──────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────┐
│ mcp_to_skill.py │
│ - Reads config │
│ - Generates Skill structure │
└──────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────┐
│ Generated Skill │
│ ├── SKILL.md (100 tokens) │
│ ├── executor.py (dynamic calls) │
│ └── config files │
└─────────────────────────────────────┘
│
▼
┌─────────────────────────────────────┐
│ Claude │
│ - Loads metadata only │
│ - Full docs when needed │
│ - Calls executor for tools │
└─────────────────────────────────────┘
# Create config
cat > github.json << 'EOF'
{
"name": "github",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {"GITHUB_TOKEN": "ghp_your_token"}
}
EOF
# Convert
python mcp_to_skill.py --mcp-config github.json --output-dir ./skills/github
# Result: GitHub tools accessible with 100 tokens vs 8k# Convert multiple MCP servers
for config in configs/*.json; do
name=$(basename "$config" .json)
python mcp_to_skill.py --mcp-config "$config" --output-dir "./skills/$name"
donepip install mcpCheck your config file:
- Command is correct
- Environment variables set
- Server is accessible
cd skills/your-skill
# List tools
python executor.py --list
# Describe a tool
python executor.py --describe tool_name
# Call a tool
python executor.py --call '{"tool": "tool_name", "arguments": {...}}'- Early stage (feedback welcome)
- Requires
mcpPython package - Some complex auth may need adjustments
- Not all MCP servers tested
This is a proof of concept. Contributions welcome:
- Test with more MCP servers
- Improve error handling
- Add more examples
- Better documentation
Inspired by:
- playwright-skill by @lackeyjb
- Anthropic Skills framework
- Model Context Protocol
MIT
Status: Functional but early stage
Feedback: Issues and PRs welcome
Questions: Open an issue